Artificial Intelligence based breast thermography using radiomic feature extraction versus conventional manual interpretation of breast thermograms in the prediction of breast cancer: a multi-reader study

ABSTRACT Objective In recent years artificial intelligence-enhanced breast thermography is increasingly being evaluated as an ancillary modality in the evaluation of breast disease. The objective of this study was to evaluate the performance of Thermalytix, a CE-marked system that analyzes thermal images using advanced thermal radiomics against unaided manual interpretation of thermographic images by trained thermologists. Methods In this retrospective, multi-reader study, thermal imaging data of 258 women who participated in a previously published clinical trial were used. These images were read manually by 3 trained thermologists independent of each other, using the approved scoring system of the American Association of Thermologists. None of the readers were involved in the collection of the images in the study cases. The images were then evaluated by the Thermalytix system, which is a commercially available software that automatically extracts hotspot, areolar and nipple radiomic parameters with a total of 64 individual radiomic features being analyzed using 3 random forest classifiers configured for 200 decision trees to generate a score predictive of the presence of breast cancer in the region of interest. The manual interpretation and Thermalytix interpretation were compared for sensitivity, specificity, positive predictive value, and negative predictive value and receiver operating characteristic curves were created to estimate prediction accuracy. Results Automated Thermalytix had sensitivity and specificity of 95.2% and 66.7% respectively while AUROC of 0.85 (13.7% greater) than manual interpretation. Further, hotspot and vascular scores derived in the automated Thermalytix are the strongest predictors of breast cancer lesions (AUROC: 0.84 and 0.83, respectively). Conclusions Overall this suggests that automated AI-based Thermalytix has higher accuracy in the prediction of breast cancer lesions and must be further investigated in the wider women population to validate its use in hospital settings as a screening modality for breast cancer.