سامانه همایش‌ها TSTA | ثبت‌نام و ارسال مقاله کنفرانس‌ها و کنگره‌های ملی

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

A Comparative Analysis of CNN Architectures for Histopathology Image Classification: Performance, Efficiency, and Adversarial Robustness

Authors :

Moein Akbari Shahpar¹ Mohsen Akbari-Shahpar²

1- Department of Engineering Science, University of Tehran, Tehran, Iran 2- Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz

Keywords :

Convolutional Neural Networks،CNN،Histopathology،Adversarial Robustness،Adversarial Training،Medical Imaging،Image Classification

Abstract :

Abstract—The integration of Convolutional Neural Networks (CNNs) into histopathology promises to revolutionize diagnostics, yet their vulnerability to adversarial attacks poses a significant risk to clinical deployment. This study investigates the relationship between CNN architecture, performance, and adversarial robustness. We benchmarked five distinct architectures (VGG16, ResNet-50, MobileNetV3-Large, EfficientNet-B4, and ConvNeXt-Tiny) on the PathMNIST dataset for classification accuracy, efficiency, and robustness against both Projected Gradient Descent (PGD) and Fast Gradient Sign Method (FGSM) attacks. Our results, averaged over five runs, show that while ConvNeXt-Tiny achieved the highest clean accuracy (93.07%±0.88%), its performance collapsed under low-strength attacks. Adversarial training significantly enhanced resilience, maintaining 62.14%±0.65% accuracy under a PGD attack that reduced the standard model's accuracy to nearly zero. This highlights a critical trade-off between standard accuracy and adversarial robustness, underscoring the need to evaluate models for both safety and reliability before clinical adoption.