Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Impact of Lesion Focus on the Performance of AI-Based Melanoma Classification

Tanay Donde

TL;DR

Melanoma detection with CNNs can be unreliable when models misfocus on non-lesion background. The authors interrogate the relationship between lesion attention and performance by using bounding boxes, masked images, and transfer learning, assessed with Grad-CAM, Sobol’, and RISE on ISIC-2019 and HAM10000. They find a positive correlation between lesion-focused attention and diagnostic metrics, with the best uplift observed when combining masked and regular images and transferring masked-trained representations, achieving an accuracy of 91.87% and $F1$ of 0.780. The work demonstrates that explainability-driven training, targeting lesion features, can improve both accuracy and interpretability, supporting the development of trustworthy AI in clinical melanoma diagnosis.

Abstract

Melanoma is the most lethal subtype of skin cancer, and early and accurate detection of this disease can greatly improve patients' outcomes. Although machine learning models, especially convolutional neural networks (CNNs), have shown great potential in automating melanoma classification, their diagnostic reliability still suffers due to inconsistent focus on lesion areas. In this study, we analyze the relationship between lesion attention and diagnostic performance, involving masked images, bounding box detection, and transfer learning. We used multiple explainability and sensitivity analysis approaches to investigate how well models aligned their attention with lesion areas and how this alignment correlated with precision, recall, and F1-score. Results showed that models with a higher focus on lesion areas achieved better diagnostic performance, suggesting the potential of interpretable AI in medical diagnostics. This study provides a foundation for developing more accurate and trustworthy melanoma classification models in the future.

The Impact of Lesion Focus on the Performance of AI-Based Melanoma Classification

TL;DR

Melanoma detection with CNNs can be unreliable when models misfocus on non-lesion background. The authors interrogate the relationship between lesion attention and performance by using bounding boxes, masked images, and transfer learning, assessed with Grad-CAM, Sobol’, and RISE on ISIC-2019 and HAM10000. They find a positive correlation between lesion-focused attention and diagnostic metrics, with the best uplift observed when combining masked and regular images and transferring masked-trained representations, achieving an accuracy of 91.87% and of 0.780. The work demonstrates that explainability-driven training, targeting lesion features, can improve both accuracy and interpretability, supporting the development of trustworthy AI in clinical melanoma diagnosis.

Abstract

Melanoma is the most lethal subtype of skin cancer, and early and accurate detection of this disease can greatly improve patients' outcomes. Although machine learning models, especially convolutional neural networks (CNNs), have shown great potential in automating melanoma classification, their diagnostic reliability still suffers due to inconsistent focus on lesion areas. In this study, we analyze the relationship between lesion attention and diagnostic performance, involving masked images, bounding box detection, and transfer learning. We used multiple explainability and sensitivity analysis approaches to investigate how well models aligned their attention with lesion areas and how this alignment correlated with precision, recall, and F1-score. Results showed that models with a higher focus on lesion areas achieved better diagnostic performance, suggesting the potential of interpretable AI in medical diagnostics. This study provides a foundation for developing more accurate and trustworthy melanoma classification models in the future.
Paper Structure (21 sections, 6 figures, 6 tables)

This paper contains 21 sections, 6 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Datasets
  5. Evaluation Metrics
  6. Baseline Model
  7. YOLO Model
  8. Masked Image Experiments
  9. Masked-to-Regular Transfer
  10. Combined Masked and Regular Datasets
  11. Results and Discussion
  12. Baseline Model Performance
  13. YOLO Model Performance
  14. Masked Image Experiment Results
  15. Masked-to-Regular Transfer Results
  16. ...and 6 more sections

Figures (6)

  • Figure 1: Grad-CAM results from the baseline InceptionV3 model, demonstrating misaligned attention away from the lesion area across various examples during melanoma classification.
  • Figure 2: Grad-CAM explanation of a correctly predicted melanoma case where the model’s attention is misaligned, focusing away from the lesion.
  • Figure 3: Representative images from each class in the ISIC-2019 dataset.
  • Figure 4: Visual examples illustrating the three attention alignment categories: fully aligned (left), partially aligned (center), and misaligned (right).
  • Figure 5: Flowchart illustrating the process of generating bounding boxes from lesion segmentation masks in the HAM10000 dataset.
  • ...and 1 more figures