SkinLesion-STN: Enhancing Skin Lesion Classification using Spatial Transformer Networks

SkinLesion-STN offers a promising solution to enhance skin lesion classification by explicitly addressing spatial variations in skin lesion images. This research paper proposes SkinLesion-STN, a novel deep model for type detection from images of skin lesions. SkinLesion-STN is a customly designed Spatial Transformer Network aimed to improve the performance of automated lesion type classifiers. Integration of STNs into the architecture contributes to improved classification accuracy and robustness, making it a valuable addition to dermatology and medical image analysis domains. The novelty of SkinLesion-STN lies in its potential to improve performance of skin lesion type detection. Thus, leading to early detection of skin cancer and reducing false negatives and false positives. The efficiency of SkinLesion-STN model is verified with extensive experiments on HAM10000 benchmark dataset. SkinLesion-STN is evaluated against EfficientNet-B0 to evaluate its performance. SkinLesion-STN has achieved good results in terms of specificity and accuracy whereas EfficientNet-B0 achieved good specificity over SkinLesion-STN. With more fine-tuned hyper parameters, SkinLesion-STN may perform better in terms of specificity also. SkinLesion-STN demonstrated an accuracy of 0.980 against EfficientNet-B0 with 0.917 accuracy. The results demonstrated superior performance, thereby establishing the superiority of STNs for skin lesion detection and classification tasks. Overall, this research is aimed to contribute to the advancement of computer-aided diagnostic systems for dermatology, ultimately benefiting patients and healthcare providers alike.