Fully Autonomous Myositis Diagnosis from Ultrasound Images Using Explainable CNN
2026 International Conference on Electrical, Computer, and Energy Technologies, ICECET 2026, Rome, İtalya, 6 - 09 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/icecet65726.2026.11632671
- Basıldığı Şehir: Rome
- Basıldığı Ülke: İtalya
- Anahtar Kelimeler: CNN, Deep learning, Grad-CAM, Medical image classification, Myositis diagnosis, Ultrasound imaging, VGG19, XAI
- İstanbul Ticaret Üniversitesi Adresli: Evet
Özet
Myositis is a rare muscle disease diagnosed using imaging techniques such as MRI and ultrasound (US). Although MRI provides reliable results, ultrasound is more practical and cost-effective. Thus, the development of automated diagnostic systems based on US images is important. In this study, we propose a fully autonomous and explainable deep learning framework for myositis diagnosis using ultrasound images. The proposed approach uses a modified VGG19 network pre-trained on ImageNet. We evaluated our model on four diagnostic scenarios involving healthy subjects and three myositis subtypes: dermatomyositis (DM), polymyositis (PM), and inclusion body myositis (IBM). Before training, images were preprocessed by removing artifacts and resizing to a smaller common resolution. The model is evaluated with 5-fold cross-validation. Experimental results show average accuracies of 94.37%, 96.39%, 93.95%, and 93.16% for S1-S4 scenarios, respectively. Our model outperforms several previous methods in most metrics while using fewer parameters. However, performance is not enough for an AI model to be used in a real hospital. AI models also need to be explainable so that real doctors can interpret them. To address the black-box nature of convolutional neural networks, we used the Grad-CAM algorithm. This will allow doctors to view the most influential parts of US images during prediction as heatmaps. Moreover, unlike some recent approaches, the proposed method does not require manual segmentation and therefore remains fully autonomous. As far as we know, this is the first explainable AI-based system for myositis diagnosis from ultrasound images.