Osteosarcoma Tumor Detection using Different Transfer Learning Models

Abstract The clinical image analysis field has increasingly embraced the utilization of transfer learning models due to their reduced computational complexity and improved accuracy, among other benefits. These models, which are pre-trained and do not require training from scratch, eliminate the need for extensive datasets. While transfer learning models are primarily employed for brain, breast, or lung image analysis, other areas such as bone marrow cell detection or bone cancer detection can also derive advantages from their use, particularly considering the limited availability of large datasets for these specific tasks. This research paper investigates the performance of various transfer learning models for the detection of osteosarcoma tumours, a form of bone cancer predominantly found in the long bones of the body's cells. The dataset comprises histopathology images stained with H\&E, categorized into four groups: Viable Tumor, Non-viable Tumor, Non-Tumor, and Viable Non-viable. The datasets were randomly split into training and test sets, following an 80-20 ratio, with 80% used for training and 20% for testing. Four models were evaluated for comparison: EfficientNetB7, InceptionResNetV2, NasNetLarge, and ResNet50. All these models were pre-trained on ImageNet. According to the results, InceptionResNetV2 achieved the highest accuracy (93.29%), followed by NasNetLarge (90.91%), ResNet50 (89.83%), and EfficientNetB7 (62.77%). InceptionResNetV2 also exhibited the highest precision (0.8658) and recall (0.8658) values among the four models.