Beyond Accuracy: A MultiDimensional Framework for Evaluating Medical Image Classification Through Win vs. Lose Model...

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning models like ResNet, optimized for natural images, often perform poorly on specialized medical imaging tasks.
• Generalization challenges arise when applying pre-trained models directly to diverse medical datasets such as skin lesion classification.

Purpose of the Study:

• To investigate the limitations of "off-the-shelf" deep learning models for skin lesion classification.
• To identify optimal deep learning architectures and resolutions for medical image analysis.
• To develop and evaluate a framework for comparing model interpretability.

Main Methods:

• Systematic evaluation of 35 architectural configurations on the DermaMNIST dataset.
• Analysis of varying image resolutions and model depths.
• Development of a cross-architectural interpretability framework using Grad-CAM heatmaps and perceptual metrics (fractal dimension, entropy, symmetry).

Main Results:

• Mid-depth architectures (3-4 layers) and intermediate resolutions (128x128) provided the best accuracy-generalization balance.
• The RevNet-layer3 model achieved superior performance (accuracy = 0.766) compared to deeper ResNet baselines.
• Fractal dimension emerged as a reliable metric for discriminating effective model attention patterns, enhancing interpretability.

Conclusions:

• Increased model depth or resolution does not inherently improve medical domain performance.
• Domain-specific architectural selection and interpretability-driven evaluation are crucial for reliable deep learning in healthcare.
• The study introduces a novel methodology for assessing performance-interpretability trade-offs in medical image classification.