대한신장학회

Skip Navigation: Skip to contents

KOR
ENG
전체메뉴보기

간행물 검색

현재 페이지 경로

HOME
간행물
간행물 검색

논문분류	춘계학술대회 초록집
제목	Automated Classification of Renal Pathologies From CT Imaging Using Deep Learning
저자	Muhammad Umair Ahmad Khan
출판정보	2026; 2026(1):
키워드	EfficientNet-B0, Renal Pathology, Computer-Aided Diagnosis , Kidney Disease Classification
초록	Objectives: Timely and accurate classification of kidney pathologies, including cysts, tumors, and stones, remains a critical clinical challenge. Manual interpretation of CT imaging is time-consuming and subject to inter-observer variability. This study presents a transfer learning framework leveraging EfficientNet-B0 to automate multi-class kidney disease detection from CT scans, aiming to support radiological workflows with a robust, generalizable diagnostic tool. Methods: A dataset of 12,446 CT images was curated across four categories: Normal, Cyst, Tumor, and Stone. Images were partitioned into training (70%), validation (15%), and test (15%) subsets. Data augmentation and ImageNet-standard normalization were applied during preprocessing. The model utilized a pretrained EfficientNet-B0 backbone, modified with a dropout layer (rate = 0.3) and a fully connected output layer (4 nodes, Softmax activation). Training was conducted in PyTorch using stochastic gradient descent (learning rate = 0.01, momentum = 0.9) over 10 epochs with a batch size of 32. Performance was evaluated via accuracy and cross-entropy loss curves. Results: The model demonstrated consistent improvement over 10 epochs, achieving a final training accuracy of approximately 97% and a test accuracy of approximately 92%. Training and test accuracy curves converged closely, with no evidence of overfitting. Test loss declined steadily from 1.2 to 0.4, confirming strong generalization. A sample inference on an unseen CT image yielded a correct prediction of "Tumor," validating the model's end-to-end utility. Conclusion: EfficientNet-B0, fine-tuned on a large-scale kidney CT dataset, achieves high multi-class classification with minimal overfitting. The model's lightweight architecture and strong generalization make it a compelling candidate for clinical deployment in resource-constrained radiology settings. Future work will focus on explainability techniques (e.g., Grad-CAM) and external dataset validation to support clinical translation.
원문(PDF)	PDF 원문보기

목록

위로가기