Sun-Exposure and Lesion Location Bias in Deep Learning Models for Skin Cancer Detection

Sun-Exposure and Lesion Location Bias in Deep Learning Models for Skin Cancer Detection | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Sun-Exposure and Lesion Location Bias in Deep Learning Models for Skin Cancer Detection Eva Milara, Vanesa Gómez-Martínez, David Chushig-Muzo, Conceição Granja, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7581023/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Deep learning (DL) models have demonstrated high performance in classifying skin lesions from dermoscopic images. However, the influence of photoexposure-related factors, such as level of exposure due to the anatomical site and skin phototype, on classification performance remains understudied. Investigating these factors is essential not only to understand their potential impact on bias in model predictions, but also to explore their potential role as risk indicators for skin cancer. Objective This study aims to assess the impact of skin phototype and anatomical-site–related photoexposure on the performance of DL models for malignancy detection, with a focus on potential sources of bias. Methods DL models are trained on widely used public dermoscopic image datasets. Performance is then evaluated on a recently published dataset of 60 patients from the University Hospital of North Norway, which includes dermoscopic images and clinical data on skin phototype and anatomical site to assess their impact on model performance. Results Preliminary analysis suggests that model performance varies between subgroups, with reduced precision observed in lesions chronically photoexposed. The reactions that cause red and painful skin are associated with better model performance, in addition to being mostly benign lesions. However, this result is skewed by the low incidence of malignant cases. Conclusions The findings highlight that individual sun-related behaviours and skin characteristics can influence the reliability of DL-based skin lesion identification. These results underscore the importance of evaluating model robustness across diverse patient profiles and may guide future efforts to define healthier sun exposure habits for the population. Bioinformatics Biomedical Engineering Oncology Artificial Intelligence and Machine Learning sun exposure skin lesions skin lesion classification image classification Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Skin cancer remains one of the most prevalent malignancies worldwide, as highlighted by the World Health Organization [ 1 ]. In 2022, global cancer statistics reported an estimated 331,722 new melanoma cases and over 1.2 million non-melanoma skin cancer diagnoses, with associated mortality rates of 58,667 and 69,416, respectively [ 2 , 3 ]. Timely identification of suspicious skin lesions is essential for improving patient outcomes, especially when dealing with malignant tumours. Traditionally, dermatological assessment has relied on expert analysis of dermoscopic images, frequently supported by manually engineered descriptors based on visual attributes such as colour, texture, or morphology [ 4 – 6 ]. In recent years, however, the field has experienced a paradigm shift with the increasing adoption of artificial intelligence (AI) tools for automatic image classification. These methods aim to promote greater diagnostic consistency and reduce inter-observer variability, thereby supporting clinicians in the early detection and classification of skin cancer [ 7 , 8 ]. While most prior research has focused on distinguishing benign from malignant lesions, relatively few studies have examined the factors underlying malignancy. Understanding these factors could inform healthier sun-exposure behaviours and guide effective prevention strategies. Of these few studies, Tognetti et al. [ 9 – 11 ] have contributed several studies focusing on the evaluation of lesion location and patient characteristics in differentiating atypical nevi from early melanoma. In a 2021 study [ 9 ], they investigated how varying degrees of photoexposure across different body regions influenced the performance of four diagnostic methods: the ABCD rule, iDScore, 7-point checklist, and visual inspection. More recently, in a 2022 study [ 10 ], they concluded that the anatomical site has a significant impact on the dermoscopic appearance of early melanomas. Specifically, lesions located on areas such as the abdomen, chest, and sides (torso) often lack conventional dermoscopic features in their initial stages, underscoring the importance of incorporating anatomical location into diagnostic considerations. Similarly, Ghiasvand et al. [ 12 ] confirm divergent pathways to melanoma development, showing that patterns of sun exposure are associated with site-specific melanoma risk. Notably, recreational sun exposure and indoor tanning were linked to melanoma on the lower limbs and torso. Despite the findings reported in these studies, the assessment has been limited to expert visual inspection or descriptors based on visual features. To the best of our knowledge, no studies have investigated the effect of sun exposure on the performance of automatic classification models based on deep learning (DL). This study aims to assess how levels of sun exposure and skin reactions impact the performance of automatic classification models for malignant skin lesion identification. To this end, four DL architectures (AlexNet [ 13 ], VGG16 [ 14 ], ResNet [ 15 ], and DenseNet [ 16 ]) are trained on three public databases (Derm7pt [ 17 ], International Skin Imaging Collaboration (ISIC)-2020 [ 18 ], and PH2 [ 19 ]). Subsequently, to measure the bias, a public dataset with information on sun exposure[ 20 ] acquired at the University Hospital of North Norway is employed. This dataset includes three skin reactions to the sun (turning brown, red, and red with pain) and four different levels of exposure based on the anatomical location of the lesion (chronically, frequently, seldom, and rarely), enabling an assessment of exposure-related bias. Methods Materials The dataset used for training the DL models consists of three public datasets, while the test dataset consists of a public dataset including photoexposure information. Both datasets are described in detail below. Training datasets In order to assess how DL models perform under different levels of sun exposure and individual skin reactions, a dataset is built by integrating images from three open-access dermoscopic repositories. Specifically, 1,011 malignancy-related cases are included from Derm7pt [ 17 ], 5,943 curated samples from ISIC-2020 [ 18 ] after discarding duplicates and ambiguous labels, and 200 high-resolution images from the PH2 dataset [ 19 ]. The available metadata varied across datasets: Derm7pt included sex, lesion location, and diagnosis; ISIC-2020 provided sex, age, lesion location, and diagnosis; and PH2 contained sex and diagnosis. However, none of them include information about sun exposure. The complete dataset, obtained by merging the three sources, was divided into 80% for training and 20% for validation, while preserving the original distribution of samples from each source and class in both subsets. Test dataset For testing the DL models, a dataset collected at the University Hospital of North Norway [ 20 ] is employed, which comprises dermoscopic images of 60 patients with sun exposure information. Ethical approval for this data collection was granted by the Regional Committee for Medical and Health Research Ethics of Northern Norway (Ref.: 392439). The dataset includes 1,518 dermoscopic images corresponding to 260 skin lesions and scars from 60 patients, of which 30 were confirmed histopathologically. The rest were validated through expert consensus among three dermatologists. Only primary lesions were considered for analysis, excluding excision scars, resulting in 216 unique lesions of 59 patients. Given that several images were acquired per lesion, a quality-ranking procedure was applied to select a single representative image per lesion. This was based on five equally weighted criteria—sharpness, noise, exposure, contrast, and blur—with features contributing positively or negatively to a cumulative score. The highest-scoring image for each lesion was selected. Different clinical and demographic data were also collected at the patient level, including variables such as age, sex, anthropometric measures, hair colour, sun sensitivity, mole characteristics, sunburn history, sunbed usage, cancer history (personal and familial), immunosuppressive status, organ transplantation, and Fitzpatrick skin type. Lesion-level metadata comprises anatomical site, diagnosis, and lesion size. Images pre-processing A consistent pre-processing workflow is applied to all images, comprising hair artifact removal, lesion segmentation, and image normalisation. Hair removal is performed using an approach that combines YCbCr colour space transformation with Attention U-Net (Att-Net)-based segmentation and image restoration through Aggregated Contextual-Transformation-Generative Adversarial Network (AOT-GAN), proposed in a previous author’s work [ 21 ]. The segmented lesion regions are obtained using the Double U-Net model [ 22 ]. Since pre-trained models are used in this work, image normalisation is conducted on a per-channel basis using the standard mean and standard deviation values from ImageNet. Sun-Exposure-Based Domain Categorisation To investigate the effects of sun exposure, different study domains are defined based on lesion location and the patient's skin response to sunlight. Regarding lesion location, four categories of photoexposure frequency are established following the classification in [ 9 , 10 ]: LA, chronically photoexposed areas, including the head, neck, and arms; LB, frequently photoexposed, covering the lower limbs up to the thighs; LC, seldom photoexposed, including the shoulders, chest, and back; and LD, rarely photoexposed, such as the abdomen, buttocks, and sides. Table 1 summarises the anatomical regions associated with each category, along with the number of lesions, the benign-to-malignant ratio, and the total number of cases per category. These categories are assigned at the lesion level, meaning that a single patient may present lesions located in different sun-exposure areas. Table 1 Definition of the four sun-exposure frequency categories used to classify lesion locations. For each category, the corresponding body areas are listed, along with the number of lesions, the proportion of benign to malignant cases, and the total number of lesions included. B: benign; M: malignant; T: total. Abb Level of exposition Anatomical location B:M (T) LA Chronically photoexposed Head/neck + Upper extremity 59:18 (77) LB Frequently photoexposed Lower extremity 8:1 (9) LC Seldom photoexposed Shoulders + Chest + Back 95:16 (111) LD Rarely photoexposed Abdomen + Bottom + Side 11:2 (13) For skin response to sun exposure, three categories are considered: RB, when the skin reacts to sun exposure by turning brown; RR, when the skin reacts by turning red; and RRp, when the skin reacts by becoming painful as well as turning red. In this case, the categories are defined at the patient level rather than the lesion level and therefore cannot be directly assigned to either benign or malignant lesions. Accordingly, among the 59 patients included in the dataset, the frequencies of RB, RR, and RRp are 22, 28, and 9, respectively. Lesion selection criteria for test set To avoid bias in the analysis of sun exposure domains, a single lesion is selected for each patient. The selection of this lesion is based on the criteria illustrated in Fig. 1 . To select one lesion per patient, the low prevalence of the LB and LD categories is considered the primary selection criterion. If multiple lesions meet this condition, the malignancy status is evaluated next, followed by the presence of histopathological confirmation. In cases where these criteria remain inconclusive, the lesion with the largest diameter is selected. It is worth noting that no patient presents lesions in both LB and LD categories. Following these criteria, Table 2 presents the final count of patients with benign and malignant lesions for each category of lesion location and skin reaction to sun exposure. It is worth highlighting the low frequency of malignant lesions in categories LB, LD, and RRp, with only 1, 2, and 1 patients, respectively. In addition, among the most frequent location categories, LA and LC, it is noteworthy that in LA, the most photoexposed areas, malignant cases are more prevalent, whereas in LC, seldom photoexposed regions, benign cases are considerably more common. Table 2 Distribution of benign and malignant lesions by location and skin reaction to sun exposure. LA: chronically photoexposed; LB: frequently photoexposed; LC: seldom photoexposed; LD: rarely photoexposed; RB: skin reacts by turning brown; RR: skin reacts by turning red; RRp: skin reacts by turning red and painful. Benign N (%) Malignant N (%) LA 8 (20.5%) 13 (65.0%) LB 6 (15.4%) 1 (5.0%) LC 18 (46.2%) 4 (20.0%) LD 7 (17.9%) 2 (10.0%) RB 13 (33.3%) 9 (45.0%) RR 18 (46.2%) 10 (50%) RRp 8 (20.5%) 1 (5.0%) Statistical analysis To evaluate the distribution of sun exposure-related domains based on lesion location and individual skin reaction in relation to malignancy, a statistical analysis is performed. Specifically, Fisher’s exact test is applied to assess associations between categorical variables, as it is particularly well-suited for small sample sizes and contingency tables with low expected frequencies [ 23 , 24 ]. To control for the increased risk of false positives due to multiple comparisons, P -values are adjusted using the Benjamini–Hochberg procedure [ 25 ]. This method is a widely accepted correction, offering greater statistical power while effectively controlling the false discovery rate, particularly in exploratory biomedical research [ 26 ]. Skin lesion classifiers For implementing skin lesion classifiers, several convolutional neural network architectures are explored, including AlexNet [ 13 ], VGG16 [ 14 ], ResNet [ 15 ], and DenseNet [ 16 ]. These models are well-established in the field of computer vision and are commonly employed in medical imaging applications [ 27 – 30 ]. For training, a five-fold strategy is adopted, where each fold involves splitting the training data into 80% for model training and 20% for validation, with each fold generated using a distinct random seed, to introduce variability. Given the class imbalance between benign and malignant cases in the training set, the training partition of each fold undergoes an undersampling process to balance the dataset. Furthermore, different hyperparameters are explored. In particular, a grid search is conducted over combinations of batch size (BS) and learning rate (LR), testing values of BS = {4, 8, 16} and LR = {10 − 5 , 10 − 4 , 10 − 3 , 10 − 2 }. Each combination is identified as Ci, with i ranging from 0 to 11, as follows: C0 (4, 10 − 5 ), C1 (4, 10 − 4 ), C2 (4, 10 − 3 ), C3 (4, 10 − 2 ), C4 (8, 10 − 5 ), C5 (8, 10 − 4 ), C6 (8, 10 − 3 ), C7 (8, 10 − 2 ), C8 (16, 10 − 5 ), C9 (16, 10 − 4 ), C10 (16, 10 − 3 ), and C11 (16, 10 − 2 ). Once all models have been trained, the test set is evaluated across all architectures and hyperparameter configurations (i.e., BS and LR combinations). The performance of each model is evaluated based on standard metrics such as accuracy, F1-score macro, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUROC) [ 31 , 32 ]. Results Statistical analysis The results of the statistical analysis using Fisher's exact test, along with the Benjamini-Hochberg correction, are presented in Table 3 . Notably, only the LA domain shows statistically significant differences between benign and malignant groups, with a P -value below 0.05 even after correction. This finding suggests that higher sun exposure may be directly associated with an increased likelihood of malignancy. Among the remaining values, all exceeding the 0.05 threshold, LC stands out with a relatively lower P -value, as expected given its high frequency. Table 3 P-values obtained from Fisher's test and their correction by means of Benjamini-Hochberg procedure. LA: chronically photoexposed; LB: frequently photoexposed; LC: seldom photoexposed; LD: rarely photoexposed; RB: skin reacts by turning brown; RR: skin reacts by turning red; RRp: skin reacts by turning red and painful. Fisher’s P -value Benjamini-Hochberg Correction P -value LA 0.0013 0.0053 LB 0.4045 0.5393 LC 0.0865 0.1730 LD 0.7040 0.7040 RB 0.4076 0.6113 RR 0.7907 0.7907 RRp 0.1480 0.4439 Classifiers evaluation To evaluate the performance of the classifiers across each domain, Table 4 presents the best-performing combinations of architecture, BS, and LR for each studied domain. Notably, the combination of VGG with a BS of 4 and an LR of 10 − 4 achieves the best results for LA, LC, RR, and RRp. This highlights it as a configuration with good generalisation capabilities for identifying malignancy in skin lesions, regardless of the characteristics related to sun exposure. Among the studied domains, an acceptable performance is first observed for LA, although it does not particularly stand out. Secondly, LB shows notably high performance, with metrics exceeding 0.85—except for sensitivity, which reaches 0.8. However, the presence of only one malignant lesion in this domain limits the representativeness of these results, as reflected in the variability of sensitivity across the five subsets. For LC, high values of sensitivity (0.9) and AUROC (0.856) are observed, in contrast with lower values in the remaining metrics, suggesting good identification of malignant cases due to their lower prevalence relative to benign ones. Regarding the domains associated with the reaction of skin to sun exposure, similar performance is observed for RB and RR, with generally acceptable metric values. In contrast, RRp stands out, with all metrics above 0.8; however, this domain includes only one malignant lesion compared to eight benign ones, which limits the interpretability of the results. Table 4 Performance metrics for different groups of patients. G: group; Arc: architecture; BS: batch size; LR: learning rate; LA: chronically photoexposed; LB: frequently photoexposed; LC: seldom photoexposed; LD: rarely photoexposed; RB: skin reacts by turning brown; RR: skin reacts by turning red; RRp: skin reacts by turning red and painful; VG: VGG; DN: DenseNet; RN: ResNet. Values greater than or equal to 0.85 are marked in bold. G Arc (BS, LR) Accuracy F1-score Sensitivity Specificity AUROC LA VG (4, 10 − 4 ) 0.686 ± 0.054 0.677 ± 0.051 0.661 ± 0.128 0.725 ± 0.105 0.679 ± 0.070 LB DN (16, 10 − 3 ) 0.917 ± 0.064 0.892 ± 0.241 0.800 ± 0.447 1.000 ± 0.000 1.000 ± 0.000 LC VG (4, 10 − 4 ) 0.627 ± 0.122 0.589 ± 0.093 0.900 ± 0.137 0.567 ± 0.173 0.856 ± 0.119 LD VG (4, 10 − 5 ) 0.889 ± 0.111 0.858 ± 0.145 0.900 ± 0.224 0.886 ± 0.120 0.972 ± 0.039 RB RN (16, 10 − 4 ) 0.700 ± 0.095 0.692 ± 0.098 0.800 ± 0.122 0.631 ± 0.213 0.713 ± 0.179 RR VG (4, 10 − 4 ) 0.686 ± 0.078 0.677 ± 0.071 0.760 ± 0.114 0.644 ± 0.145 0.735 ± 0.040 RRp VG (4, 10 − 4 ) 0.867 ± 0.122 0.807 ± 0.176 1.000 ± 0.000 0.850 ± 0.137 0.975 ± 0.056 Level of exposition domains To assess the variability in model performance across location domains, Fig. 2 is presented. This figure highlights both the low variability and moderate performance of LA, in contrast with the other domains, with LB being the most noteworthy. In this case, the performance of AlexNet is highly variable, limiting its practical utility. However, most BS and LR combinations in VGG and ResNet achieve outstanding performance values (between 0.8 and 1). For LC, lower performance with limited variability is observed, whereas LD exhibits higher overall performance but greater variability. In LC, the VGG architecture stands out for its superior performance and stability, while in LD this trend is observed in ResNet. To further assess model performance, Fig. 3 displays the AUROC values for each combination of BS and LR across the different architectures and location-related domains. First, AlexNet (AN) exhibits high variability across configurations, with inconsistent performance between domains. Notably, although a few combinations (e.g., C0 and C8) yield outstanding AUROC values (close to 1.0) for LB and LD, many others remain acceptable (below 0.6), especially for LC and LA. This suggests that AlexNet lacks robustness and generalisation capacity under this multi-domain setup. VGG (VG) demonstrates the most consistent and high-performing behaviour across all domains. Most configurations yield high AUROC values (above 0.8), especially in LD (e.g., C5: 0.83, C6: 0.86, C8: 0.94, C9: 0.94). This supports the idea that VGG is a suitable architecture for skin lesion classification regardless of sun exposure domain, with strong performance particularly in less frequently photoexposed regions like LD. In the case of ResNet (RN), performance is slightly less stable than VGG but still robust. For LB, multiple configurations (e.g., C9 and C10) reach excellent AUROC values (0.93, 0.97 or even 1.0), highlighting its strength in frequently photoexposed regions. LC and LD show a drop in AUROC, possibly indicating difficulty in distinguishing benign from malignant lesions in those domains with this architecture, being better for LD than LC. DenseNet (DN) achieves high AUROC values for LB across nearly all configurations (near 1.0). However, performance in LC and LD is more modest, with lower values and greater variability (ranging from ~ 0.47 to 0.84). This suggests DenseNet may be more effective in handling lesions in frequently photoexposed areas, but less so in rarely exposed regions. Furthermore, all combinations of these four architectures perform poorly for LA, with regions chronically exposed to light being the most difficult to classify. Overall, VGG appears to offer the best trade-off between generalisation and robustness across exposure domains and training configurations. In contrast, the performance of AlexNet is highly configuration-sensitive, limiting its practical utility in clinical applications requiring domain-invariant prediction. ResNet and DenseNet perform well in selected domains but are slightly more domain-dependent than VGG. Skin reaction to sun exposure domains As with lesion location, the variability in model performance across domains defined by skin reaction to sun exposure is assessed using boxplots for each architecture (see Fig. 4 ). In this case, a similar behaviour is observed for RB and RR, the most frequent domains, characterised by lower overall performance but reduced variability. For RB, no specific architecture stands out. In contrast, for RR, DenseNet shows a more promising performance despite higher variability, reaching the highest AUROC (up to 0.7). For the RRp domain, most architectures achieve substantially higher AUROC values (between 0.6 and 1), except AlexNet. However, the limited number of malignant cases within this domain (only one) constrains the interpretability of these results. To further assess the impact of each combination of architecture, BS, and LR across domains, Fig. 5 presents a heatmap of the corresponding AUROC values. Notably, AlexNet consistently shows poor performance across all domains. In contrast, for the remaining architectures, certain combinations—such as C0 or C9—stand out, reaching high AUROC values for RB and RR (around 0.7), and exceeding for RRp (0.95). Regarding this latter domain, VGG, ResNet, and DenseNet demonstrate generally strong performance, with several BS and LR combinations yielding the highest AUROC values (above or close to 0.8). From a clinical perspective, it is worth noting that the RRp domain represents patients with the most severe reaction to sun exposure, as it involves both sunburn and associated pain. However, this group predominantly comprises benign cases (8 out of 9), in contrast to the RB group, which reflects a milder skin response to sunlight but presents a more balanced distribution of benign and malignant lesions. Discussion The identification of risk factors and healthy sun exposure habits remains essential for reducing the incidence of skin cancer. Some studies have addressed these aspects through expert assessments [ 9 – 12 ]. Nonetheless, to the best of our knowledge, none have integrated them into automatic classification models. This study aims to explore this integration by evaluating the impact of sun exposure and the reaction of skin to it on the performance of DL classifiers for malignant lesions identification. To this end, a public database containing information on both the anatomical location of the lesion, which enables the definition of the level of sun exposure, and the type of skin reaction to such exposure is employed. It is noteworthy that lesions from the most sun-exposed group (LA) are predominantly malignant, while benign lesions are more common in the other exposure levels. With respect to skin reaction types, all are characterised by predominantly benign lesions. For the individual evaluation of each exposure level and skin reaction type, the statistical analysis confirms a strong association between the highest exposure level (LA) and malignant lesions, consistent with the incidence observed. Clinically, this finding aligns with the well-established role of chronic sun exposure as a major risk factor for skin cancer. However, for the remaining variables, no statistically significant results are obtained, which limits their interpretability and suggests that their potential role may be less pronounced or requires larger cohorts to be properly assessed. The evaluation of classifier performance across domains reveals clear relationships between model architecture, domain characteristics, and clinical implications. For levels of exposition to sun domains, VGG with a BS of 4 and a LR of 10 − 4 consistently achieves the best results across LA, LC, RR, and RRp. Clinically, this indicates that VGG can reliably identify malignancy regardless of whether lesions occur in frequently or rarely sun-exposed regions, making it suitable for general use across diverse anatomical locations. In the LA domain, acceptable performance is observed, though metrics are lower than in other levels of exposition. Sensitivity remains modest, reflecting difficulties in correctly identifying malignant lesions in chronically sun-exposed areas. This aligns with clinical expectations, as long-term sun exposure can produce benign skin changes that mimic malignancy, increasing diagnostic complexity. In contrast, LB exhibits high overall metrics (all above 0.85 except for sensitivity at 0.8), suggesting models can effectively identify malignancy in frequently photoexposed regions. However, the presence of only one malignant lesion in the test dataset limits the generalizability of this finding, highlighting the need for more balanced datasets. For LC, sensitivity (0.9) and AUROC (0.856) indicate strong identification of malignant cases despite lower prevalence. Clinically, this suggests that lesions in the level of exposition are less likely to be overlooked by DL models, but other metrics indicate that benign lesions may still occasionally be misclassified. LD shows higher performance overall but with greater variability, particularly benefiting from ResNet configurations. This reflects the potential for accurate malignancy detection in rarely exposed areas, though results may fluctuate depending on the training setup. Regarding skin reaction to sun exposure, RB and RR demonstrate moderate performance with reduced variability. These domains correspond to milder photoreactions, suggesting that the classifiers perform consistently when lesions are associated with common skin responses. In RRp, all metrics exceed 0.8 for VGG, ResNet, and DenseNet, yet the domain included only one malignant lesion among eight benign cases. Clinically, this indicates that while DL can accurately classify severe photoreaction cases, the predominance of benign lesions limits confidence in malignant detection. Across all domains, AlexNet shows high variability and generally poor robustness, emphasising that architecture choice is critical for reliable classification. DenseNet achieves excellent AUROC values in frequently photoexposed regions like LB, but it is less effective in rarely exposed areas such as LC and LD, suggesting domain-dependent strengths. VGG offers the most stable performance across all domains, making it the preferred architecture for clinical applications where domain-invariant predictions are required. In summary, these results indicate that automated skin lesion classification benefits from architectures capable of generalisation across diverse anatomical locations and sun exposure profiles. Domains with chronically exposed skin or limited malignant samples remain challenging, reflecting real-world clinical complexity. Future work should aim to expand datasets across all exposure and photoreaction domains, ensuring that models can accurately detect malignancy even in rare or atypical presentations. Despite the novelty of this study, several limitations must be acknowledged. The evaluation relies on a highly restricted dataset of only 60 patients, limiting the analysis to 59 images when assessing the biological factors under consideration, and most lesions lack histopathological confirmation of malignancy. Within this dataset, the number of cases per category is particularly small, especially for malignant lesions across most domains. For sun exposure levels, the LB and LD groups are markedly underrepresented, with very few malignant cases also observed in LB, LC, and LD. Similarly, for skin reaction types, the dataset includes only a small number of patients exhibiting red skin accompanied by pain. Moreover, lesion selection per patient is guided by observations from the dataset—such as the low incidence of LB and LD lesions and the reduced prevalence of malignant cases—which may have introduced a benign-to-malignant ratio that does not faithfully reflect the original distribution. Finally, with respect to the implemented methodology, further limitations should be acknowledged. Only four DL architectures (AlexNet, VGG, ResNet, and DenseNet) are evaluated, despite the wide range of more recent architectures that have demonstrated superior efficiency in this context. In addition, hyperparameters are explored exclusively within a fixed grid, rather than being optimised through adaptive or automated search strategies. Considering both the results and the identified limitations, future research should primarily focus on the development of larger and more diverse datasets that systematically incorporate information on additional exposure-related factors that could not be assessed in the present study (e.g., use of sun protection, artificial tanning devices, sunbathing patterns). With such expanded and heterogeneous datasets, it would be possible to more accurately characterise the risk factors for skin cancer associated with sun exposure. Conclusions This study demonstrates that the performance of automatic classifiers for skin lesion diagnosis is influenced by both the level of photoexposure based on lesion location and skin reaction to sun exposure, suggesting that these domains may reflect underlying biological risk factors. The lower performance in chronically sun-exposed regions (LA) indicates that cumulative photodamage generates benign patterns that resemble malignant ones, making both clinical and automated diagnosis more challenging. Conversely, cases with severe acute photoreactions (RRp) were predominantly benign, suggesting that this phenotype is not necessarily associated with higher malignant risk. These results highlight the potential of DL not only as a tool that supports clinicians in diagnosis but also as a means to investigate how sun exposure patterns and skin responses are related to malignancy risk. Future research should prioritise the development of larger and more balanced datasets to refine the identification of such risk factors, ultimately improving both clinical prevention strategies and automated diagnostic accuracy. Abbreviations AI Artificial Intelligence AN AlexNet AOT-GAN Aggregated Contextual-Transformation-Generative Adversarial Network Att-Net Attention U-Net AUROC Area under the receiver operating characteristic curve BS Batch size DL Deep learning DN DenseNet ISIC International Skin Imaging Collaboration LA Chronically photoexposed LB Frequently photoexposed LC Seldom photoexposed LD Rarely photoexposed LR Learning rate RB Skin reacts by turning brown RN ResNet RR Skin reacts by turning red RRp Skin reacts by turning red and painful VG VGG WARIFA Watching the risk factors:Artificial intelligence and the prevention of chronic conditions Declarations Conflicts of Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements This work was supported by the European Commission through the H2020-EU.3.1.4.2, European Project WARIFA (Watching the risk factors: Artificial intelligence and the prevention of chronic conditions) under Grant Agreement 101017385; and by the Spanish federal grants PID2019-107768RA-I00 \& PID2023-149457OB-I00 (all funded by the agency AEI/ 10.13039/501100011033 ). The study sponsors have not been involved in any stage of the study. References International Agency for Research on Cancer, Skin cancer – IARC, (2025). https://www.iarc.who.int/ (accessed September 4, 2025). J. Ferlay, M. Ervik, F. Lam, M. Laversanne, M. Colombet, L. Mery, M. Piñeros, A. Znaor, I. Soerjomataram, F. Bray, Global Cancer Observatory: Cancer Today, (2024). H. Sung, J. Ferlay, R.L. Siegel, M. Laversanne, I. Soerjomataram, A. Jemal, F. Bray, Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries, CA Cancer J Clin (2021). https://doi.org/10.3322/caac.21660. T.J. Brinker, A. Hekler, A.H. Enk, J. Klode, A. Hauschild, C. Berking, B. Schilling, S. Haferkamp, D. Schadendorf, C. von Kalle, Deep learning outperformed 136 of 157 dermatologists in a head-to-head dermoscopic melanoma image classification task, Eur J Cancer 113 (2019) 47–54. https://doi.org/10.1016/j.ejca.2019.04.001. C. Barata, M.E. Celebi, J.S. Marques, A survey of feature extraction in dermoscopy image analysis of skin cancer, IEEE J Biomed Health Inform 23 (2018) 1096–1109. https://doi.org/10.1109/JBHI.2018.2845939. V. Gómez-Martínez, D. Chushig-Muzo, M.B. Veierød, C. Granja, C. Soguero-Ruiz, Ensemble feature selection and tabular data augmentation with generative adversarial networks to enhance cutaneous melanoma identification and interpretability, BioData Min 17 (2024) 46. https://doi.org/10.1186/s13040-024-00397-7. U.A. Lyakhova, P.A. Lyakhov, Systematic review of approaches to detection and classification of skin cancer using artificial intelligence: Development and prospects, Comput Biol Med 178 (2024) 108742. https://doi.org/10.1016/j.compbiomed.2024.108742. F. Grignaffini, F. Barbuto, L. Piazzo, M. Troiano, P. Simeoni, F. Mangini, G. Pellacani, C. Cantisani, F. Frezza, Machine Learning Approaches for Skin Cancer Classification from Dermoscopic Images: A Systematic Review, Algorithms 15 (2022) 438. https://doi.org/10.3390/a15110438. L. Tognetti, A. Cartocci, E. Cinotti, E. Moscarella, F. Farnetani, A. Lallas, D. Tiodorovic, C. Carrera, C. Longo, S. Puig, J. ~L. Perrot, G. Argenziano, G. Pellacani, G. Cataldo, A. Balistreri, G. Cevenini, P. Rubegni, The impact of anatomical location and sun exposure on the dermoscopic recognition of atypical nevi and early melanomas: usefulness of an integrated clinical-dermoscopic method (iDScore), Journal of the European Academy of Dermatology and Venereology 35 (2021) 650–657. https://doi.org/10.1111/jdv.16847. L. Tognetti, A. Cartocci, E. Cinotti, E. Moscarella, F. Farnetani, C. Carrera, A. Lallas, D. Tiodorovic, C. Longo, S. Puig, J.L. Perrot, G. Argenziano, G. Pellacani, G. Cataldo, A. Balistreri, G. Cevenini, P. Rubegni, Dermoscopy of early melanomas: variation according to the anatomic site, Arch Dermatol Res 314 (2022) 183–190. https://doi.org/10.1007/s00403-021-02226-x. L. Tognetti, A. Cartocci, E. Cinotti, M. D’Onghia, M. Zychowska, E. Moscarella, E. Dika, F. Farnetani, S. Guida, J. Paoli, A. Lallas, D. Tiodorovic, I. Stanganelli, C. Longo, M. Suppa, I. Zalaudek, G. Argenziano, J.L. Perrot, G. Rubegni, G. Cataldo, P. Rubegni, Dermoscopy of atypical pigmented lesions of the face: Variation according to facial areas, Exp Dermatol 32 (2023) 2166–2172. https://doi.org/10.1111/exd.14941. R. Ghiasvand, T.E. Robsahm, A.C. Green, C.S. Rueegg, E. Weiderpass, E. Lund, M.B. Veierød, Association of Phenotypic Characteristics and UV Radiation Exposure With Risk of Melanoma on Different Body Sites, JAMA Dermatol 155 (2019) 39–49. https://doi.org/10.1001/jamadermatol.2018.3964. A. Krizhevsky, I. Sutskever, G.E. Hinton, ImageNet classification with deep convolutional neural networks, in: Advances in Neural Information Processing Systems (NeurIPS), 2012: pp. 1097–1105. K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, in: International Conference on Learning Representations (ICLR), 2015. K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016: pp. 770–778. G. Huang, Z. Liu, L. Van Der Maaten, K.Q. Weinberger, Densely connected convolutional networks, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017: pp. 4700–4708. J. Kawahara, S. Daneshvar, G. Argenziano, G. Hamarneh, Seven-Point Checklist and Skin Lesion Classification Using Multitask Multimodal Neural Nets, IEEE J Biomed Health Inform 23 (2019) 538–546. https://doi.org/10.1109/JBHI.2018.2824327. B. Cassidy, C. Kendrick, A. Brodzicki, J. Jaworek-Korjakowska, M.H. Yap, Analysis of the ISIC image datasets: Usage, benchmarks and recommendations, Med Image Anal 75 (2022). https://doi.org/10.1016/j.media.2021.102305. teresa Mendoça, P.M. Ferreira, A.R.S. Marçal, C. Barata, J.S. Marques, J. Rocha, J. Rozeira, Accurate and Scalable System for Automatic Detection of Malignant Melanoma, in: Dermoscopy Image Analysis, CRC Press, 2015: pp. 309–360. https://doi.org/10.1201/b19107-14. M. Castro-Fernandez, T. Schopf, I. Castaño-Gonzalez, B. Roque, H. Kirchesch, S. Ortega Sarmiento, H. Fabelo, F. Godtliebsen, C. Granja, G. Callico, MCR-SL: A Multimodal, Context-Rich Skin Lesion Dataset for Skin Cancer Diagnosis, (2025). https://doi.org/10.5281/zenodo.17056062. V. Gómez-Martínez, others, A Data-Driven Approach for Digital Hair Removal in Dermoscopy Images Using Encoder-Decoder and Generative Adversarial Network-Based Models, 2024. D. Jha, M.A. Riegler, D. Johansen, P. Halvorsen, H.D. Johansen, DoubleU-Net: A deep convolutional neural network for medical image segmentation, in: Proceedings of the IEEE Symposium on Computer-Based Medical Systems (CBMS), Institute of Electrical and Electronics Engineers Inc., 2020: pp. 558–564. R.A. Fisher, On the interpretation of χ^2 from contingency tables, and the calculation of P, Journal of the Royal Statistical Society 85 (1922) 87–94. J.H. McDonald, Handbook of Biological Statistics (2nd ed.), (2009). http://www.biostathandbook.com. Y. Benjamini, Y. Hochberg, Controlling the false discovery rate: a practical and powerful approach to multiple testing, Journal of the Royal Statistical Society. Series B (Methodological) 57 (1995) 289–300. M.E. Glickman, S.R. Rao, M.R. Schultz, False discovery rate control is a recommended alternative to Bonferroni-type adjustments in health studies, J Clin Epidemiol 67 (2014) 850–857. Y. Wu, B. Chen, A. Zeng, D. Pan, R. Wang, S. Zhao, Skin Cancer Classification With Deep Learning: A Systematic Review, Front Oncol 12 (2022) 893972. https://doi.org/10.3389/fonc.2022.893972. D. Popescu, M. El-Khatib, H. El-Khatib, L. Ichim, New Trends in Melanoma Detection Using Neural Networks: A Systematic Review, Sensors 22 (2022) 496. https://doi.org/10.3390/s22020496. A. Naeem, M.S. Farooq, A. Khelifi, A. Abid, Malignant Melanoma Classification Using Deep Learning: Datasets, Performance Measurements, Challenges and Opportunities, IEEE Access 8 (2020) 110575–110597. https://doi.org/10.1109/ACCESS.2020.3001507. J. Saeed, S. Zeebaree, Skin Lesion Classification Based on Deep Convolutional Neural Networks Architectures, Journal of Applied Science and Technology Trends 2 (2021) 41–51. https://doi.org/10.38094/jastt20189. M. Sokolova, G. Lapalme, Performance measures in classification of human communications, Inf Process Manag 45 (2009) 427–437. T. Fawcett, An introduction to ROC analysis, Pattern Recognit Lett 27 (2006) 861–874. Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7581023","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":512994839,"identity":"70bb1635-7afa-45c4-8066-d35a1c49881b","order_by":0,"name":"Eva Milara","email":"","orcid":"https://orcid.org/0000-0002-6955-7312","institution":"Universidad Rey Juan Carlos","correspondingAuthor":false,"prefix":"","firstName":"Eva","middleName":"","lastName":"Milara","suffix":""},{"id":512994840,"identity":"caa36acf-2abe-4bfc-8d4c-d9ffe6138bd1","order_by":1,"name":"Vanesa Gómez-Martínez","email":"","orcid":"https://orcid.org/0009-0001-3349-3900","institution":"Universidad Rey Juan Carlos","correspondingAuthor":false,"prefix":"","firstName":"Vanesa","middleName":"","lastName":"Gómez-Martínez","suffix":""},{"id":512994841,"identity":"38614e28-22af-45c5-8d9c-1d2333052eb1","order_by":2,"name":"David Chushig-Muzo","email":"","orcid":"https://orcid.org/0000-0001-5585-2305","institution":"Universidad Rey Juan Carlos","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"","lastName":"Chushig-Muzo","suffix":""},{"id":512994842,"identity":"f61228e2-a17e-4eea-b618-25de804b3de5","order_by":3,"name":"Conceição Granja","email":"","orcid":"https://orcid.org/0000-0002-3028-8899","institution":"University Hospital of North Norway","correspondingAuthor":false,"prefix":"","firstName":"Conceição","middleName":"","lastName":"Granja","suffix":""},{"id":512994843,"identity":"553c6c78-fc4a-4042-9dc1-e7c39cdd090f","order_by":4,"name":"Cristina Soguero-Ruiz","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDklEQVRIiWNgGAWjYFACxgYGBgMkPj8DKh+7lgPISiQbCGoBggPIHIMD2BUhnDH7cPPnDwV35Bn4Dx9+zVNxT874Ru6Dgh8MNva4tEicS2yTOGDwzLBBIi3NmudMsbHZjXQDwx6GtMQGHFoMeBjbgH45zNggwWNmzNuWkLjtRhqDMQPD4QRctgC1NH8AarFv4D8D1PIvIXHzDLCW/zgdBtTSAHTYYaAzcowf8zYkJG6QAGs5wIjLYRJnGNskzhgcTm4D+oVxzrEEY4kzzxgMewyScfqFv4f98YeKP4dt+4Eh9uFNTYIcf3sam8GPCjucDoMDNiCSgLENiIhMMGD+AGM8IE7DKBgFo2AUjBAAAKD2UrlmQaiuAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0001-5817-989X","institution":"Universidad Rey Juan Carlos","correspondingAuthor":true,"prefix":"","firstName":"Cristina","middleName":"","lastName":"Soguero-Ruiz","suffix":""}],"badges":[],"createdAt":"2025-09-10 09:12:06","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7581023/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7581023/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":91078965,"identity":"1d55f5a7-4fa4-4d01-a6bf-df0092677bf4","added_by":"auto","created_at":"2025-09-11 11:21:54","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":82054,"visible":true,"origin":"","legend":"\u003cp\u003eLesion selection criteria for the test set based on location categories, malignancy, and histopathological confirmation. LA: chronically photoexposed; LB: frequently photoexposed; LC: seldom photoexposed; LD: rarely photoexposed.\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7581023/v1/17d1ec806a18a8b565214548.png"},{"id":91078964,"identity":"f86ad3d7-93dc-46bc-a4db-01e4efef9d96","added_by":"auto","created_at":"2025-09-11 11:21:54","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":43809,"visible":true,"origin":"","legend":"\u003cp\u003eDifference in AUROC performance across lesion locations. LA: chronically photoexposed; LB: frequently photoexposed; LC: seldom photoexposed; LD: rarely photoexposed.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7581023/v1/ffd8f3419eaece3fa6884ccd.png"},{"id":91078969,"identity":"5dd55610-71f0-411e-940c-6f903d3da812","added_by":"auto","created_at":"2025-09-11 11:21:54","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":224049,"visible":true,"origin":"","legend":"\u003cp\u003eDifference in AUROC performance across lesion locations. AN: AlexNet; VG: VGG; RN: ResNet; and DN: DenseNet; LA: chronically photoexposed; LB: frequently photoexposed; LC: seldom photoexposed; LD: rarely photoexposed.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7581023/v1/2eac540b6a69daef709b04b2.png"},{"id":91080077,"identity":"2c5c7175-3fdc-47ec-9740-8c9a4739d122","added_by":"auto","created_at":"2025-09-11 11:29:54","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":44433,"visible":true,"origin":"","legend":"\u003cp\u003eDifference in AUROC performance across skin reactions to the sun. RB: skin reacts by turning brown; RR: skin reacts by turning red; RRp: skin reacts by turning red and painful.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7581023/v1/6277e56282956733027324ea.png"},{"id":91078974,"identity":"8b409d8c-ca76-4d46-b83a-eaa60e0802a2","added_by":"auto","created_at":"2025-09-11 11:21:55","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":183687,"visible":true,"origin":"","legend":"\u003cp\u003eDifference in AUROC performance across skin reactions to the sun. AN: AlexNet; VG: VGG; RN: ResNet; and DN: DenseNet; RB: skin reacts by turning brown; RR: skin reacts by turning red; RRp: skin reacts by turning red and painful.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7581023/v1/4f32fa57bd83dd270b1af7cb.png"},{"id":91082678,"identity":"10019602-a1ae-4380-903d-0f7bed31d6c7","added_by":"auto","created_at":"2025-09-11 11:53:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1286961,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7581023/v1/79bf15bd-ccc1-4a6b-a8b6-7d3d361e00b2.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eSun-Exposure and Lesion Location Bias in Deep Learning Models for Skin Cancer Detection\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSkin cancer remains one of the most prevalent malignancies worldwide, as highlighted by the World Health Organization [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. In 2022, global cancer statistics reported an estimated 331,722 new melanoma cases and over 1.2\u0026nbsp;million non-melanoma skin cancer diagnoses, with associated mortality rates of 58,667 and 69,416, respectively [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Timely identification of suspicious skin lesions is essential for improving patient outcomes, especially when dealing with malignant tumours. Traditionally, dermatological assessment has relied on expert analysis of dermoscopic images, frequently supported by manually engineered descriptors based on visual attributes such as colour, texture, or morphology [\u003cspan additionalcitationids=\"CR5\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. In recent years, however, the field has experienced a paradigm shift with the increasing adoption of artificial intelligence (AI) tools for automatic image classification. These methods aim to promote greater diagnostic consistency and reduce inter-observer variability, thereby supporting clinicians in the early detection and classification of skin cancer [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eWhile most prior research has focused on distinguishing benign from malignant lesions, relatively few studies have examined the factors underlying malignancy. Understanding these factors could inform healthier sun-exposure behaviours and guide effective prevention strategies. Of these few studies, Tognetti \u003cem\u003eet al.\u003c/em\u003e [\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] have contributed several studies focusing on the evaluation of lesion location and patient characteristics in differentiating atypical nevi from early melanoma. In a 2021 study [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], they investigated how varying degrees of photoexposure across different body regions influenced the performance of four diagnostic methods: the ABCD rule, iDScore, 7-point checklist, and visual inspection. More recently, in a 2022 study [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], they concluded that the anatomical site has a significant impact on the dermoscopic appearance of early melanomas. Specifically, lesions located on areas such as the abdomen, chest, and sides (torso) often lack conventional dermoscopic features in their initial stages, underscoring the importance of incorporating anatomical location into diagnostic considerations. Similarly, Ghiasvand \u003cem\u003eet al.\u003c/em\u003e [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] confirm divergent pathways to melanoma development, showing that patterns of sun exposure are associated with site-specific melanoma risk. Notably, recreational sun exposure and indoor tanning were linked to melanoma on the lower limbs and torso. Despite the findings reported in these studies, the assessment has been limited to expert visual inspection or descriptors based on visual features. To the best of our knowledge, no studies have investigated the effect of sun exposure on the performance of automatic classification models based on deep learning (DL).\u003c/p\u003e\u003cp\u003eThis study aims to assess how levels of sun exposure and skin reactions impact the performance of automatic classification models for malignant skin lesion identification. To this end, four DL architectures (AlexNet [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], VGG16 [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], ResNet [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], and DenseNet [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]) are trained on three public databases (Derm7pt [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], International Skin Imaging Collaboration (ISIC)-2020 [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], and PH2 [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]). Subsequently, to measure the bias, a public dataset with information on sun exposure[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] acquired at the University Hospital of North Norway is employed. This dataset includes three skin reactions to the sun (turning brown, red, and red with pain) and four different levels of exposure based on the anatomical location of the lesion (chronically, frequently, seldom, and rarely), enabling an assessment of exposure-related bias.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eMaterials\u003c/h2\u003e\u003cp\u003eThe dataset used for training the DL models consists of three public datasets, while the test dataset consists of a public dataset including photoexposure information. Both datasets are described in detail below.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eTraining datasets\u003c/h3\u003e\n\u003cp\u003eIn order to assess how DL models perform under different levels of sun exposure and individual skin reactions, a dataset is built by integrating images from three open-access dermoscopic repositories. Specifically, 1,011 malignancy-related cases are included from Derm7pt [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], 5,943 curated samples from ISIC-2020 [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] after discarding duplicates and ambiguous labels, and 200 high-resolution images from the PH2 dataset [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. The available metadata varied across datasets: Derm7pt included sex, lesion location, and diagnosis; ISIC-2020 provided sex, age, lesion location, and diagnosis; and PH2 contained sex and diagnosis. However, none of them include information about sun exposure. The complete dataset, obtained by merging the three sources, was divided into 80% for training and 20% for validation, while preserving the original distribution of samples from each source and class in both subsets.\u003c/p\u003e\n\u003ch3\u003eTest dataset\u003c/h3\u003e\n\u003cp\u003eFor testing the DL models, a dataset collected at the University Hospital of North Norway [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] is employed, which comprises dermoscopic images of 60 patients with sun exposure information. Ethical approval for this data collection was granted by the Regional Committee for Medical and Health Research Ethics of Northern Norway (Ref.: 392439).\u003c/p\u003e\u003cp\u003eThe dataset includes 1,518 dermoscopic images corresponding to 260 skin lesions and scars from 60 patients, of which 30 were confirmed histopathologically. The rest were validated through expert consensus among three dermatologists. Only primary lesions were considered for analysis, excluding excision scars, resulting in 216 unique lesions of 59 patients. Given that several images were acquired per lesion, a quality-ranking procedure was applied to select a single representative image per lesion. This was based on five equally weighted criteria\u0026mdash;sharpness, noise, exposure, contrast, and blur\u0026mdash;with features contributing positively or negatively to a cumulative score. The highest-scoring image for each lesion was selected.\u003c/p\u003e\u003cp\u003eDifferent clinical and demographic data were also collected at the patient level, including variables such as age, sex, anthropometric measures, hair colour, sun sensitivity, mole characteristics, sunburn history, sunbed usage, cancer history (personal and familial), immunosuppressive status, organ transplantation, and Fitzpatrick skin type. Lesion-level metadata comprises anatomical site, diagnosis, and lesion size.\u003c/p\u003e\n\u003ch3\u003eImages pre-processing\u003c/h3\u003e\n\u003cp\u003eA consistent pre-processing workflow is applied to all images, comprising hair artifact removal, lesion segmentation, and image normalisation. Hair removal is performed using an approach that combines YCbCr colour space transformation with Attention U-Net (Att-Net)-based segmentation and image restoration through Aggregated Contextual-Transformation-Generative Adversarial Network (AOT-GAN), proposed in a previous author\u0026rsquo;s work [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. The segmented lesion regions are obtained using the Double U-Net model [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Since pre-trained models are used in this work, image normalisation is conducted on a per-channel basis using the standard mean and standard deviation values from ImageNet.\u003c/p\u003e\n\u003ch3\u003eSun-Exposure-Based Domain Categorisation\u003c/h3\u003e\n\u003cp\u003eTo investigate the effects of sun exposure, different study domains are defined based on lesion location and the patient's skin response to sunlight. Regarding lesion location, four categories of photoexposure frequency are established following the classification in [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]: LA, chronically photoexposed areas, including the head, neck, and arms; LB, frequently photoexposed, covering the lower limbs up to the thighs; LC, seldom photoexposed, including the shoulders, chest, and back; and LD, rarely photoexposed, such as the abdomen, buttocks, and sides. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarises the anatomical regions associated with each category, along with the number of lesions, the benign-to-malignant ratio, and the total number of cases per category. These categories are assigned at the lesion level, meaning that a single patient may present lesions located in different sun-exposure areas.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eDefinition of the four sun-exposure frequency categories used to classify lesion locations. For each category, the corresponding body areas are listed, along with the number of lesions, the proportion of benign to malignant cases, and the total number of lesions included. B: benign; M: malignant; T: total.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAbb\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLevel of exposition\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAnatomical location\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eB:M (T)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eChronically photoexposed\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eHead/neck\u0026thinsp;+\u0026thinsp;Upper extremity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e59:18 (77)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFrequently photoexposed\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLower extremity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e8:1 (9)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSeldom photoexposed\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eShoulders\u0026thinsp;+\u0026thinsp;Chest\u0026thinsp;+\u0026thinsp;Back\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e95:16 (111)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLD\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRarely photoexposed\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAbdomen\u0026thinsp;+\u0026thinsp;Bottom\u0026thinsp;+\u0026thinsp;Side\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e11:2 (13)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eFor skin response to sun exposure, three categories are considered: RB, when the skin reacts to sun exposure by turning brown; RR, when the skin reacts by turning red; and RRp, when the skin reacts by becoming painful as well as turning red. In this case, the categories are defined at the patient level rather than the lesion level and therefore cannot be directly assigned to either benign or malignant lesions. Accordingly, among the 59 patients included in the dataset, the frequencies of RB, RR, and RRp are 22, 28, and 9, respectively.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eLesion selection criteria for test set\u003c/h2\u003e\u003cp\u003eTo avoid bias in the analysis of sun exposure domains, a single lesion is selected for each patient. The selection of this lesion is based on the criteria illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. To select one lesion per patient, the low prevalence of the LB and LD categories is considered the primary selection criterion. If multiple lesions meet this condition, the malignancy status is evaluated next, followed by the presence of histopathological confirmation. In cases where these criteria remain inconclusive, the lesion with the largest diameter is selected. It is worth noting that no patient presents lesions in both LB and LD categories.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFollowing these criteria, Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents the final count of patients with benign and malignant lesions for each category of lesion location and skin reaction to sun exposure. It is worth highlighting the low frequency of malignant lesions in categories LB, LD, and RRp, with only 1, 2, and 1 patients, respectively. In addition, among the most frequent location categories, LA and LC, it is noteworthy that in LA, the most photoexposed areas, malignant cases are more prevalent, whereas in LC, seldom photoexposed regions, benign cases are considerably more common.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eDistribution of benign and malignant lesions by location and skin reaction to sun exposure. LA: chronically photoexposed; LB: frequently photoexposed; LC: seldom photoexposed; LD: rarely photoexposed; RB: skin reacts by turning brown; RR: skin reacts by turning red; RRp: skin reacts by turning red and painful.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eBenign \u003c/p\u003e\u003cp\u003eN (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMalignant\u003c/p\u003e\u003cp\u003eN (%)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e8 (20.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e13 (65.0%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e6 (15.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1 (5.0%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e18 (46.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e4 (20.0%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLD\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e7 (17.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2 (10.0%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e13 (33.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e9 (45.0%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e18 (46.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e10 (50%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRRp\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e8 (20.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1 (5.0%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003eStatistical analysis\u003c/h2\u003e\u003cp\u003eTo evaluate the distribution of sun exposure-related domains based on lesion location and individual skin reaction in relation to malignancy, a statistical analysis is performed. Specifically, Fisher\u0026rsquo;s exact test is applied to assess associations between categorical variables, as it is particularly well-suited for small sample sizes and contingency tables with low expected frequencies [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. To control for the increased risk of false positives due to multiple comparisons, \u003cem\u003eP\u003c/em\u003e-values are adjusted using the Benjamini\u0026ndash;Hochberg procedure [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. This method is a widely accepted correction, offering greater statistical power while effectively controlling the false discovery rate, particularly in exploratory biomedical research [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eSkin lesion classifiers\u003c/h3\u003e\n\u003cp\u003eFor implementing skin lesion classifiers, several convolutional neural network architectures are explored, including AlexNet [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], VGG16 [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], ResNet [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], and DenseNet [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. These models are well-established in the field of computer vision and are commonly employed in medical imaging applications [\u003cspan additionalcitationids=\"CR28 CR29\" citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. For training, a five-fold strategy is adopted, where each fold involves splitting the training data into 80% for model training and 20% for validation, with each fold generated using a distinct random seed, to introduce variability. Given the class imbalance between benign and malignant cases in the training set, the training partition of each fold undergoes an undersampling process to balance the dataset. Furthermore, different hyperparameters are explored. In particular, a grid search is conducted over combinations of batch size (BS) and learning rate (LR), testing values of BS = {4, 8, 16} and LR = {10\u003csup\u003e\u0026minus;\u0026thinsp;5\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;2\u003c/sup\u003e}. Each combination is identified as Ci, with i ranging from 0 to 11, as follows: C0 (4, 10\u003csup\u003e\u0026minus;\u0026thinsp;5\u003c/sup\u003e), C1 (4, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e), C2 (4, 10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e), C3 (4, 10\u003csup\u003e\u0026minus;\u0026thinsp;2\u003c/sup\u003e), C4 (8, 10\u003csup\u003e\u0026minus;\u0026thinsp;5\u003c/sup\u003e), C5 (8, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e), C6 (8, 10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e), C7 (8, 10\u003csup\u003e\u0026minus;\u0026thinsp;2\u003c/sup\u003e), C8 (16, 10\u003csup\u003e\u0026minus;\u0026thinsp;5\u003c/sup\u003e), C9 (16, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e), C10 (16, 10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e), and C11 (16, 10\u003csup\u003e\u0026minus;\u0026thinsp;2\u003c/sup\u003e).\u003c/p\u003e\u003cp\u003eOnce all models have been trained, the test set is evaluated across all architectures and hyperparameter configurations (i.e., BS and LR combinations). The performance of each model is evaluated based on standard metrics such as accuracy, F1-score macro, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUROC) [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eStatistical analysis\u003c/h2\u003e\u003cp\u003eThe results of the statistical analysis using Fisher's exact test, along with the Benjamini-Hochberg correction, are presented in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Notably, only the LA domain shows statistically significant differences between benign and malignant groups, with a \u003cem\u003eP\u003c/em\u003e-value below 0.05 even after correction. This finding suggests that higher sun exposure may be directly associated with an increased likelihood of malignancy. Among the remaining values, all exceeding the 0.05 threshold, LC stands out with a relatively lower \u003cem\u003eP\u003c/em\u003e-value, as expected given its high frequency.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eP-values obtained from Fisher's test and their correction by means of Benjamini-Hochberg procedure. LA: chronically photoexposed; LB: frequently photoexposed; LC: seldom photoexposed; LD: rarely photoexposed; RB: skin reacts by turning brown; RR: skin reacts by turning red; RRp: skin reacts by turning red and painful.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFisher\u0026rsquo;s\u003c/p\u003e\u003cp\u003e\u003cem\u003eP\u003c/em\u003e-value\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eBenjamini-Hochberg\u003c/p\u003e\u003cp\u003eCorrection \u003cem\u003eP\u003c/em\u003e-value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.0013\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.0053\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.4045\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.5393\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.0865\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.1730\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLD\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.7040\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.7040\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.4076\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.6113\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.7907\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.7907\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRRp\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.1480\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.4439\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eClassifiers evaluation\u003c/h2\u003e\u003cp\u003eTo evaluate the performance of the classifiers across each domain, Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents the best-performing combinations of architecture, BS, and LR for each studied domain. Notably, the combination of VGG with a BS of 4 and an LR of 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e achieves the best results for LA, LC, RR, and RRp. This highlights it as a configuration with good generalisation capabilities for identifying malignancy in skin lesions, regardless of the characteristics related to sun exposure. Among the studied domains, an acceptable performance is first observed for LA, although it does not particularly stand out. Secondly, LB shows notably high performance, with metrics exceeding 0.85\u0026mdash;except for sensitivity, which reaches 0.8. However, the presence of only one malignant lesion in this domain limits the representativeness of these results, as reflected in the variability of sensitivity across the five subsets. For LC, high values of sensitivity (0.9) and AUROC (0.856) are observed, in contrast with lower values in the remaining metrics, suggesting good identification of malignant cases due to their lower prevalence relative to benign ones. Regarding the domains associated with the reaction of skin to sun exposure, similar performance is observed for RB and RR, with generally acceptable metric values. In contrast, RRp stands out, with all metrics above 0.8; however, this domain includes only one malignant lesion compared to eight benign ones, which limits the interpretability of the results.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance metrics for different groups of patients. G: group; Arc: architecture; BS: batch size; LR: learning rate; LA: chronically photoexposed; LB: frequently photoexposed; LC: seldom photoexposed; LD: rarely photoexposed; RB: skin reacts by turning brown; RR: skin reacts by turning red; RRp: skin reacts by turning red and painful; VG: VGG; DN: DenseNet; RN: ResNet. Values greater than or equal to 0.85 are marked in bold.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eG\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eArc (BS, LR)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eF1-score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eAUROC\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eVG (4, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.686\u0026thinsp;\u0026plusmn;\u0026thinsp;0.054\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.677\u0026thinsp;\u0026plusmn;\u0026thinsp;0.051\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.661\u0026thinsp;\u0026plusmn;\u0026thinsp;0.128\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.725\u0026thinsp;\u0026plusmn;\u0026thinsp;0.105\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.679\u0026thinsp;\u0026plusmn;\u0026thinsp;0.070\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDN (16, 10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.917\u0026thinsp;\u0026plusmn;\u0026thinsp;0.064\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.892\u0026thinsp;\u0026plusmn;\u0026thinsp;0.241\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.800\u0026thinsp;\u0026plusmn;\u0026thinsp;0.447\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e1.000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.000\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e1.000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.000\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eVG (4, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.627\u0026thinsp;\u0026plusmn;\u0026thinsp;0.122\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.589\u0026thinsp;\u0026plusmn;\u0026thinsp;0.093\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.900\u0026thinsp;\u0026plusmn;\u0026thinsp;0.137\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.567\u0026thinsp;\u0026plusmn;\u0026thinsp;0.173\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.856\u0026thinsp;\u0026plusmn;\u0026thinsp;0.119\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLD\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eVG (4, 10\u003csup\u003e\u0026minus;\u0026thinsp;5\u003c/sup\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.889\u0026thinsp;\u0026plusmn;\u0026thinsp;0.111\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.858\u0026thinsp;\u0026plusmn;\u0026thinsp;0.145\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.900\u0026thinsp;\u0026plusmn;\u0026thinsp;0.224\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.886\u0026thinsp;\u0026plusmn;\u0026thinsp;0.120\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.972\u0026thinsp;\u0026plusmn;\u0026thinsp;0.039\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRN (16, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.700\u0026thinsp;\u0026plusmn;\u0026thinsp;0.095\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.692\u0026thinsp;\u0026plusmn;\u0026thinsp;0.098\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.800\u0026thinsp;\u0026plusmn;\u0026thinsp;0.122\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.631\u0026thinsp;\u0026plusmn;\u0026thinsp;0.213\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.713\u0026thinsp;\u0026plusmn;\u0026thinsp;0.179\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eVG (4, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e0.686\u0026thinsp;\u0026plusmn;\u0026thinsp;0.078\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.677\u0026thinsp;\u0026plusmn;\u0026thinsp;0.071\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e0.760\u0026thinsp;\u0026plusmn;\u0026thinsp;0.114\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e0.644\u0026thinsp;\u0026plusmn;\u0026thinsp;0.145\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e0.735\u0026thinsp;\u0026plusmn;\u0026thinsp;0.040\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRRp\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eVG (4, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.867\u0026thinsp;\u0026plusmn;\u0026thinsp;0.122\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e\u003cp\u003e0.807\u0026thinsp;\u0026plusmn;\u0026thinsp;0.176\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e\u003cp\u003e1.000\u0026thinsp;\u0026plusmn;\u0026thinsp;0.000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.850\u0026thinsp;\u0026plusmn;\u0026thinsp;0.137\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.975\u0026thinsp;\u0026plusmn;\u0026thinsp;0.056\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003eLevel of exposition domains\u003c/h2\u003e\u003cp\u003eTo assess the variability in model performance across location domains, Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e is presented. This figure highlights both the low variability and moderate performance of LA, in contrast with the other domains, with LB being the most noteworthy. In this case, the performance of AlexNet is highly variable, limiting its practical utility. However, most BS and LR combinations in VGG and ResNet achieve outstanding performance values (between 0.8 and 1). For LC, lower performance with limited variability is observed, whereas LD exhibits higher overall performance but greater variability. In LC, the VGG architecture stands out for its superior performance and stability, while in LD this trend is observed in ResNet.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo further assess model performance, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e displays the AUROC values for each combination of BS and LR across the different architectures and location-related domains. First, AlexNet (AN) exhibits high variability across configurations, with inconsistent performance between domains. Notably, although a few combinations (e.g., C0 and C8) yield outstanding AUROC values (close to 1.0) for LB and LD, many others remain acceptable (below 0.6), especially for LC and LA. This suggests that AlexNet lacks robustness and generalisation capacity under this multi-domain setup. VGG (VG) demonstrates the most consistent and high-performing behaviour across all domains. Most configurations yield high AUROC values (above 0.8), especially in LD (e.g., C5: 0.83, C6: 0.86, C8: 0.94, C9: 0.94). This supports the idea that VGG is a suitable architecture for skin lesion classification regardless of sun exposure domain, with strong performance particularly in less frequently photoexposed regions like LD. In the case of ResNet (RN), performance is slightly less stable than VGG but still robust. For LB, multiple configurations (e.g., C9 and C10) reach excellent AUROC values (0.93, 0.97 or even 1.0), highlighting its strength in frequently photoexposed regions. LC and LD show a drop in AUROC, possibly indicating difficulty in distinguishing benign from malignant lesions in those domains with this architecture, being better for LD than LC. DenseNet (DN) achieves high AUROC values for LB across nearly all configurations (near 1.0). However, performance in LC and LD is more modest, with lower values and greater variability (ranging from ~\u0026thinsp;0.47 to 0.84). This suggests DenseNet may be more effective in handling lesions in frequently photoexposed areas, but less so in rarely exposed regions. Furthermore, all combinations of these four architectures perform poorly for LA, with regions chronically exposed to light being the most difficult to classify. Overall, VGG appears to offer the best trade-off between generalisation and robustness across exposure domains and training configurations. In contrast, the performance of AlexNet is highly configuration-sensitive, limiting its practical utility in clinical applications requiring domain-invariant prediction. ResNet and DenseNet perform well in selected domains but are slightly more domain-dependent than VGG.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eSkin reaction to sun exposure domains\u003c/h2\u003e\u003cp\u003eAs with lesion location, the variability in model performance across domains defined by skin reaction to sun exposure is assessed using boxplots for each architecture (see Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). In this case, a similar behaviour is observed for RB and RR, the most frequent domains, characterised by lower overall performance but reduced variability. For RB, no specific architecture stands out. In contrast, for RR, DenseNet shows a more promising performance despite higher variability, reaching the highest AUROC (up to 0.7). For the RRp domain, most architectures achieve substantially higher AUROC values (between 0.6 and 1), except AlexNet. However, the limited number of malignant cases within this domain (only one) constrains the interpretability of these results.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo further assess the impact of each combination of architecture, BS, and LR across domains, Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e presents a heatmap of the corresponding AUROC values. Notably, AlexNet consistently shows poor performance across all domains. In contrast, for the remaining architectures, certain combinations\u0026mdash;such as C0 or C9\u0026mdash;stand out, reaching high AUROC values for RB and RR (around 0.7), and exceeding for RRp (0.95). Regarding this latter domain, VGG, ResNet, and DenseNet demonstrate generally strong performance, with several BS and LR combinations yielding the highest AUROC values (above or close to 0.8). From a clinical perspective, it is worth noting that the RRp domain represents patients with the most severe reaction to sun exposure, as it involves both sunburn and associated pain. However, this group predominantly comprises benign cases (8 out of 9), in contrast to the RB group, which reflects a milder skin response to sunlight but presents a more balanced distribution of benign and malignant lesions.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe identification of risk factors and healthy sun exposure habits remains essential for reducing the incidence of skin cancer. Some studies have addressed these aspects through expert assessments [\u003cspan additionalcitationids=\"CR10 CR11\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Nonetheless, to the best of our knowledge, none have integrated them into automatic classification models. This study aims to explore this integration by evaluating the impact of sun exposure and the reaction of skin to it on the performance of DL classifiers for malignant lesions identification.\u003c/p\u003e\u003cp\u003eTo this end, a public database containing information on both the anatomical location of the lesion, which enables the definition of the level of sun exposure, and the type of skin reaction to such exposure is employed. It is noteworthy that lesions from the most sun-exposed group (LA) are predominantly malignant, while benign lesions are more common in the other exposure levels. With respect to skin reaction types, all are characterised by predominantly benign lesions. For the individual evaluation of each exposure level and skin reaction type, the statistical analysis confirms a strong association between the highest exposure level (LA) and malignant lesions, consistent with the incidence observed. Clinically, this finding aligns with the well-established role of chronic sun exposure as a major risk factor for skin cancer. However, for the remaining variables, no statistically significant results are obtained, which limits their interpretability and suggests that their potential role may be less pronounced or requires larger cohorts to be properly assessed.\u003c/p\u003e\u003cp\u003eThe evaluation of classifier performance across domains reveals clear relationships between model architecture, domain characteristics, and clinical implications. For levels of exposition to sun domains, VGG with a BS of 4 and a LR of 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e consistently achieves the best results across LA, LC, RR, and RRp. Clinically, this indicates that VGG can reliably identify malignancy regardless of whether lesions occur in frequently or rarely sun-exposed regions, making it suitable for general use across diverse anatomical locations. In the LA domain, acceptable performance is observed, though metrics are lower than in other levels of exposition. Sensitivity remains modest, reflecting difficulties in correctly identifying malignant lesions in chronically sun-exposed areas. This aligns with clinical expectations, as long-term sun exposure can produce benign skin changes that mimic malignancy, increasing diagnostic complexity. In contrast, LB exhibits high overall metrics (all above 0.85 except for sensitivity at 0.8), suggesting models can effectively identify malignancy in frequently photoexposed regions. However, the presence of only one malignant lesion in the test dataset limits the generalizability of this finding, highlighting the need for more balanced datasets. For LC, sensitivity (0.9) and AUROC (0.856) indicate strong identification of malignant cases despite lower prevalence. Clinically, this suggests that lesions in the level of exposition are less likely to be overlooked by DL models, but other metrics indicate that benign lesions may still occasionally be misclassified. LD shows higher performance overall but with greater variability, particularly benefiting from ResNet configurations. This reflects the potential for accurate malignancy detection in rarely exposed areas, though results may fluctuate depending on the training setup.\u003c/p\u003e\u003cp\u003eRegarding skin reaction to sun exposure, RB and RR demonstrate moderate performance with reduced variability. These domains correspond to milder photoreactions, suggesting that the classifiers perform consistently when lesions are associated with common skin responses. In RRp, all metrics exceed 0.8 for VGG, ResNet, and DenseNet, yet the domain included only one malignant lesion among eight benign cases. Clinically, this indicates that while DL can accurately classify severe photoreaction cases, the predominance of benign lesions limits confidence in malignant detection.\u003c/p\u003e\u003cp\u003eAcross all domains, AlexNet shows high variability and generally poor robustness, emphasising that architecture choice is critical for reliable classification. DenseNet achieves excellent AUROC values in frequently photoexposed regions like LB, but it is less effective in rarely exposed areas such as LC and LD, suggesting domain-dependent strengths. VGG offers the most stable performance across all domains, making it the preferred architecture for clinical applications where domain-invariant predictions are required.\u003c/p\u003e\u003cp\u003eIn summary, these results indicate that automated skin lesion classification benefits from architectures capable of generalisation across diverse anatomical locations and sun exposure profiles. Domains with chronically exposed skin or limited malignant samples remain challenging, reflecting real-world clinical complexity. Future work should aim to expand datasets across all exposure and photoreaction domains, ensuring that models can accurately detect malignancy even in rare or atypical presentations.\u003c/p\u003e\u003cp\u003eDespite the novelty of this study, several limitations must be acknowledged. The evaluation relies on a highly restricted dataset of only 60 patients, limiting the analysis to 59 images when assessing the biological factors under consideration, and most lesions lack histopathological confirmation of malignancy. Within this dataset, the number of cases per category is particularly small, especially for malignant lesions across most domains. For sun exposure levels, the LB and LD groups are markedly underrepresented, with very few malignant cases also observed in LB, LC, and LD. Similarly, for skin reaction types, the dataset includes only a small number of patients exhibiting red skin accompanied by pain. Moreover, lesion selection per patient is guided by observations from the dataset\u0026mdash;such as the low incidence of LB and LD lesions and the reduced prevalence of malignant cases\u0026mdash;which may have introduced a benign-to-malignant ratio that does not faithfully reflect the original distribution. Finally, with respect to the implemented methodology, further limitations should be acknowledged. Only four DL architectures (AlexNet, VGG, ResNet, and DenseNet) are evaluated, despite the wide range of more recent architectures that have demonstrated superior efficiency in this context. In addition, hyperparameters are explored exclusively within a fixed grid, rather than being optimised through adaptive or automated search strategies.\u003c/p\u003e\u003cp\u003eConsidering both the results and the identified limitations, future research should primarily focus on the development of larger and more diverse datasets that systematically incorporate information on additional exposure-related factors that could not be assessed in the present study (e.g., use of sun protection, artificial tanning devices, sunbathing patterns). With such expanded and heterogeneous datasets, it would be possible to more accurately characterise the risk factors for skin cancer associated with sun exposure.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThis study demonstrates that the performance of automatic classifiers for skin lesion diagnosis is influenced by both the level of photoexposure based on lesion location and skin reaction to sun exposure, suggesting that these domains may reflect underlying biological risk factors. The lower performance in chronically sun-exposed regions (LA) indicates that cumulative photodamage generates benign patterns that resemble malignant ones, making both clinical and automated diagnosis more challenging. Conversely, cases with severe acute photoreactions (RRp) were predominantly benign, suggesting that this phenotype is not necessarily associated with higher malignant risk. These results highlight the potential of DL not only as a tool that supports clinicians in diagnosis but also as a means to investigate how sun exposure patterns and skin responses are related to malignancy risk. Future research should prioritise the development of larger and more balanced datasets to refine the identification of such risk factors, ultimately improving both clinical prevention strategies and automated diagnostic accuracy.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eAI\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eArtificial Intelligence\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eAN\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eAlexNet\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eAOT-GAN\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eAggregated Contextual-Transformation-Generative Adversarial Network\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eAtt-Net\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eAttention U-Net\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eAUROC\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eArea under the receiver operating characteristic curve\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eBS\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eBatch size\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eDL\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eDeep learning\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eDN\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eDenseNet\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eISIC\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eInternational Skin Imaging Collaboration\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eLA\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eChronically photoexposed\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eLB\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eFrequently photoexposed\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eLC\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eSeldom photoexposed\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eLD\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eRarely photoexposed\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eLR\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eLearning rate\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eRB\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eSkin reacts by turning brown\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eRN\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eResNet\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eRR\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eSkin reacts by turning red\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eRRp\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eSkin reacts by turning red and painful\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eVG\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eVGG\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eWARIFA\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eWatching the risk factors:Artificial intelligence and the prevention of chronic conditions\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eConflicts of Interest\u003c/h2\u003e\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e\u003cp\u003eThis work was supported by the European Commission through the H2020-EU.3.1.4.2, European Project WARIFA (Watching the risk factors: Artificial intelligence and the prevention of chronic conditions) under Grant Agreement 101017385; and by the Spanish federal grants PID2019-107768RA-I00 \\\u0026amp; PID2023-149457OB-I00 (all funded by the agency AEI/\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.13039/501100011033\u003c/span\u003e\u003cspan address=\"10.13039/501100011033\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The study sponsors have not been involved in any stage of the study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eInternational Agency for Research on Cancer, Skin cancer \u0026ndash; IARC, (2025). https://www.iarc.who.int/ (accessed September 4, 2025).\u003c/li\u003e\n \u003cli\u003eJ. Ferlay, M. Ervik, F. Lam, M. Laversanne, M. Colombet, L. Mery, M. Pi\u0026ntilde;eros, A. Znaor, I. Soerjomataram, F. Bray, Global Cancer Observatory: Cancer Today, (2024).\u003c/li\u003e\n \u003cli\u003eH. Sung, J. Ferlay, R.L. Siegel, M. Laversanne, I. Soerjomataram, A. Jemal, F. Bray, Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries, CA Cancer J Clin (2021). https://doi.org/10.3322/caac.21660.\u003c/li\u003e\n \u003cli\u003eT.J. Brinker, A. Hekler, A.H. Enk, J. Klode, A. Hauschild, C. Berking, B. Schilling, S. Haferkamp, D. Schadendorf, C. von Kalle, Deep learning outperformed 136 of 157 dermatologists in a head-to-head dermoscopic melanoma image classification task, Eur J Cancer 113 (2019) 47\u0026ndash;54. https://doi.org/10.1016/j.ejca.2019.04.001.\u003c/li\u003e\n \u003cli\u003eC. Barata, M.E. Celebi, J.S. Marques, A survey of feature extraction in dermoscopy image analysis of skin cancer, IEEE J Biomed Health Inform 23 (2018) 1096\u0026ndash;1109. https://doi.org/10.1109/JBHI.2018.2845939.\u003c/li\u003e\n \u003cli\u003eV. G\u0026oacute;mez-Mart\u0026iacute;nez, D. Chushig-Muzo, M.B. Veier\u0026oslash;d, C. Granja, C. Soguero-Ruiz, Ensemble feature selection and tabular data augmentation with generative adversarial networks to enhance cutaneous melanoma identification and interpretability, BioData Min 17 (2024) 46. https://doi.org/10.1186/s13040-024-00397-7.\u003c/li\u003e\n \u003cli\u003eU.A. Lyakhova, P.A. Lyakhov, Systematic review of approaches to detection and classification of skin cancer using artificial intelligence: Development and prospects, Comput Biol Med 178 (2024) 108742. https://doi.org/10.1016/j.compbiomed.2024.108742.\u003c/li\u003e\n \u003cli\u003eF. Grignaffini, F. Barbuto, L. Piazzo, M. Troiano, P. Simeoni, F. Mangini, G. Pellacani, C. Cantisani, F. Frezza, Machine Learning Approaches for Skin Cancer Classification from Dermoscopic Images: A Systematic Review, Algorithms 15 (2022) 438. https://doi.org/10.3390/a15110438.\u003c/li\u003e\n \u003cli\u003eL. Tognetti, A. Cartocci, E. Cinotti, E. Moscarella, F. Farnetani, A. Lallas, D. Tiodorovic, C. Carrera, C. Longo, S. Puig, J. ~L. Perrot, G. Argenziano, G. Pellacani, G. Cataldo, A. Balistreri, G. Cevenini, P. Rubegni, The impact of anatomical location and sun exposure on the dermoscopic recognition of atypical nevi and early melanomas: usefulness of an integrated clinical-dermoscopic method (iDScore), Journal of the European Academy of Dermatology and Venereology 35 (2021) 650\u0026ndash;657. https://doi.org/10.1111/jdv.16847.\u003c/li\u003e\n \u003cli\u003eL. Tognetti, A. Cartocci, E. Cinotti, E. Moscarella, F. Farnetani, C. Carrera, A. Lallas, D. Tiodorovic, C. Longo, S. Puig, J.L. Perrot, G. Argenziano, G. Pellacani, G. Cataldo, A. Balistreri, G. Cevenini, P. Rubegni, Dermoscopy of early melanomas: variation according to the anatomic site, Arch Dermatol Res 314 (2022) 183\u0026ndash;190. https://doi.org/10.1007/s00403-021-02226-x.\u003c/li\u003e\n \u003cli\u003eL. Tognetti, A. Cartocci, E. Cinotti, M. D\u0026rsquo;Onghia, M. Zychowska, E. Moscarella, E. Dika, F. Farnetani, S. Guida, J. Paoli, A. Lallas, D. Tiodorovic, I. Stanganelli, C. Longo, M. Suppa, I. Zalaudek, G. Argenziano, J.L. Perrot, G. Rubegni, G. Cataldo, P. Rubegni, Dermoscopy of atypical pigmented lesions of the face: Variation according to facial areas, Exp Dermatol 32 (2023) 2166\u0026ndash;2172. https://doi.org/10.1111/exd.14941.\u003c/li\u003e\n \u003cli\u003eR. Ghiasvand, T.E. Robsahm, A.C. Green, C.S. Rueegg, E. Weiderpass, E. Lund, M.B. Veier\u0026oslash;d, Association of Phenotypic Characteristics and UV Radiation Exposure With Risk of Melanoma on Different Body Sites, JAMA Dermatol 155 (2019) 39\u0026ndash;49. https://doi.org/10.1001/jamadermatol.2018.3964.\u003c/li\u003e\n \u003cli\u003eA. Krizhevsky, I. Sutskever, G.E. Hinton, ImageNet classification with deep convolutional neural networks, in: Advances in Neural Information Processing Systems (NeurIPS), 2012: pp. 1097\u0026ndash;1105.\u003c/li\u003e\n \u003cli\u003eK. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, in: International Conference on Learning Representations (ICLR), 2015.\u003c/li\u003e\n \u003cli\u003eK. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016: pp. 770\u0026ndash;778.\u003c/li\u003e\n \u003cli\u003eG. Huang, Z. Liu, L. Van Der Maaten, K.Q. Weinberger, Densely connected convolutional networks, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017: pp. 4700\u0026ndash;4708.\u003c/li\u003e\n \u003cli\u003eJ. Kawahara, S. Daneshvar, G. Argenziano, G. Hamarneh, Seven-Point Checklist and Skin Lesion Classification Using Multitask Multimodal Neural Nets, IEEE J Biomed Health Inform 23 (2019) 538\u0026ndash;546. https://doi.org/10.1109/JBHI.2018.2824327.\u003c/li\u003e\n \u003cli\u003eB. Cassidy, C. Kendrick, A. Brodzicki, J. Jaworek-Korjakowska, M.H. Yap, Analysis of the ISIC image datasets: Usage, benchmarks and recommendations, Med Image Anal 75 (2022). https://doi.org/10.1016/j.media.2021.102305.\u003c/li\u003e\n \u003cli\u003eteresa Mendo\u0026ccedil;a, P.M. Ferreira, A.R.S. Mar\u0026ccedil;al, C. Barata, J.S. Marques, J. Rocha, J. Rozeira, Accurate and Scalable System for Automatic Detection of Malignant Melanoma, in: Dermoscopy Image Analysis, CRC Press, 2015: pp. 309\u0026ndash;360. https://doi.org/10.1201/b19107-14.\u003c/li\u003e\n \u003cli\u003eM. Castro-Fernandez, T. Schopf, I. Casta\u0026ntilde;o-Gonzalez, B. Roque, H. Kirchesch, S. Ortega Sarmiento, H. Fabelo, F. Godtliebsen, C. Granja, G. Callico, MCR-SL: A Multimodal, Context-Rich Skin Lesion Dataset for Skin Cancer Diagnosis, (2025). https://doi.org/10.5281/zenodo.17056062.\u003c/li\u003e\n \u003cli\u003eV. G\u0026oacute;mez-Mart\u0026iacute;nez, others, A Data-Driven Approach for Digital Hair Removal in Dermoscopy Images Using Encoder-Decoder and Generative Adversarial Network-Based Models, 2024.\u003c/li\u003e\n \u003cli\u003eD. Jha, M.A. Riegler, D. Johansen, P. Halvorsen, H.D. Johansen, DoubleU-Net: A deep convolutional neural network for medical image segmentation, in: Proceedings of the IEEE Symposium on Computer-Based Medical Systems (CBMS), Institute of Electrical and Electronics Engineers Inc., 2020: pp. 558\u0026ndash;564.\u003c/li\u003e\n \u003cli\u003eR.A. Fisher, On the interpretation of \u0026chi;^2 from contingency tables, and the calculation of P, Journal of the Royal Statistical Society 85 (1922) 87\u0026ndash;94.\u003c/li\u003e\n \u003cli\u003eJ.H. McDonald, Handbook of Biological Statistics (2nd ed.), (2009). http://www.biostathandbook.com.\u003c/li\u003e\n \u003cli\u003eY. Benjamini, Y. Hochberg, Controlling the false discovery rate: a practical and powerful approach to multiple testing, Journal of the Royal Statistical Society. Series B (Methodological) 57 (1995) 289\u0026ndash;300.\u003c/li\u003e\n \u003cli\u003eM.E. Glickman, S.R. Rao, M.R. Schultz, False discovery rate control is a recommended alternative to Bonferroni-type adjustments in health studies, J Clin Epidemiol 67 (2014) 850\u0026ndash;857.\u003c/li\u003e\n \u003cli\u003eY. Wu, B. Chen, A. Zeng, D. Pan, R. Wang, S. Zhao, Skin Cancer Classification With Deep Learning: A Systematic Review, Front Oncol 12 (2022) 893972. https://doi.org/10.3389/fonc.2022.893972.\u003c/li\u003e\n \u003cli\u003eD. Popescu, M. El-Khatib, H. El-Khatib, L. Ichim, New Trends in Melanoma Detection Using Neural Networks: A Systematic Review, Sensors 22 (2022) 496. https://doi.org/10.3390/s22020496.\u003c/li\u003e\n \u003cli\u003eA. Naeem, M.S. Farooq, A. Khelifi, A. Abid, Malignant Melanoma Classification Using Deep Learning: Datasets, Performance Measurements, Challenges and Opportunities, IEEE Access 8 (2020) 110575\u0026ndash;110597. https://doi.org/10.1109/ACCESS.2020.3001507.\u003c/li\u003e\n \u003cli\u003eJ. Saeed, S. Zeebaree, Skin Lesion Classification Based on Deep Convolutional Neural Networks Architectures, Journal of Applied Science and Technology Trends 2 (2021) 41\u0026ndash;51. https://doi.org/10.38094/jastt20189.\u003c/li\u003e\n \u003cli\u003eM. Sokolova, G. Lapalme, Performance measures in classification of human communications, Inf Process Manag 45 (2009) 427\u0026ndash;437.\u003c/li\u003e\n \u003cli\u003eT. Fawcett, An introduction to ROC analysis, Pattern Recognit Lett 27 (2006) 861\u0026ndash;874.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[{"identity":"9c061630-9337-4328-b908-42256865fd5f","identifier":"10.13039/501100000780","name":"European Commission","awardNumber":"101017385","order_by":0}],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"King Juan Carlos University","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"sun exposure, skin lesions, skin lesion classification, image classification","lastPublishedDoi":"10.21203/rs.3.rs-7581023/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7581023/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eDeep learning (DL) models have demonstrated high performance in classifying skin lesions from dermoscopic images. However, the influence of photoexposure-related factors, such as level of exposure due to the anatomical site and skin phototype, on classification performance remains understudied. Investigating these factors is essential not only to understand their potential impact on bias in model predictions, but also to explore their potential role as risk indicators for skin cancer.\u003c/p\u003e\u003ch2\u003eObjective\u003c/h2\u003e\u003cp\u003eThis study aims to assess the impact of skin phototype and anatomical-site\u0026ndash;related photoexposure on the performance of DL models for malignancy detection, with a focus on potential sources of bias.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eDL models are trained on widely used public dermoscopic image datasets. Performance is then evaluated on a recently published dataset of 60 patients from the University Hospital of North Norway, which includes dermoscopic images and clinical data on skin phototype and anatomical site to assess their impact on model performance.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003ePreliminary analysis suggests that model performance varies between subgroups, with reduced precision observed in lesions chronically photoexposed. The reactions that cause red and painful skin are associated with better model performance, in addition to being mostly benign lesions. However, this result is skewed by the low incidence of malignant cases.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eThe findings highlight that individual sun-related behaviours and skin characteristics can influence the reliability of DL-based skin lesion identification. These results underscore the importance of evaluating model robustness across diverse patient profiles and may guide future efforts to define healthier sun exposure habits for the population.\u003c/p\u003e","manuscriptTitle":"Sun-Exposure and Lesion Location Bias in Deep Learning Models for Skin Cancer Detection","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-11 11:21:50","doi":"10.21203/rs.3.rs-7581023/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"13214b23-5fce-4d6b-bdd3-f5ac1da9e94c","owner":[],"postedDate":"September 11th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":54495846,"name":"Bioinformatics"},{"id":54495847,"name":"Biomedical Engineering"},{"id":54495848,"name":"Oncology"},{"id":54495849,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-09-11T11:21:50+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-11 11:21:50","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7581023","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7581023","identity":"rs-7581023","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}