سامانه همایش های دانشگاه آزاد اسلامی واحد نجف آباد

صفحه اصلی / اولین همایش بین المللی تحول خدمات سلامت: تلفیق مراقبت، درمان، بازتوانی و امیدبخشی

Deep Learning-Based Multi-Cancer Detection from Heterogeneous Medical Imaging: A Unified Convolutional Neural Network Framework for Early Diagnosis and Clinical Management

نویسندگان :

Sajjad Ghatei¹ Mohammad Rouhi² Zohreh Arabi³

1- دانشگاه ملی مهارت 2- دانشگاه پیام نور 3- دانشگاه ملی مهارت

کلمات کلیدی :

deep learning،cancer detection،convolutional neural network،multi-task learning،transfer learning،transfer learning،medical image analysis،Grad-CAM

چکیده :

Background: Cancer is responsible for nearly 10 million deaths each year and is the second leading cause of mortality worldwide. Breast, lung, colorectal, and skin cancers together cause more than 4.3 million deaths annually. Since early diagnosis greatly increases survival rates, limited access to imaging interpretation remains a major challenge, especially given the WHO estimate of a global shortage of 1.3 million radiologists, with greater impact in low-resource regions. Method: A unified multi-task convolutional neural network based on a shared ResNet-50 architecture, enhanced with modality-specific normalisation layers and four dedicated sigmoid classification outputs, was developed. The model was trained using 218,450 labelled images from mammography, low-dose CT, colonoscopy, and dermoscopy datasets. Adaptive uncertainty-based loss weighting was applied to minimise conflicts among tasks. Evaluation used AUC, sensitivity, specificity, F1-score, and comparison against 12 specialist radiologists and four single-task baselines. Finding: Evaluation on 25,438 test images produced a mean AUC of 0.968 (breast: 0.981, lung: 0.974, colorectal: 0.962, skin: 0.956), outperforming the radiologist panel (mean AUC 0.956; breast cancer p=0.012). Multi-task learning improved mean AUC by 3.0 percentage points over single-task baselines. Grad-CAM plausibility was confirmed in 92.3% of true-positive cases. The system reduced parameter count by 62% and completed four-modality analysis in under two seconds. Conclusion: The unified multi-task CNN provides radiologist-level multi-cancer detection with superior efficiency and scalability compared to single-task models, supporting deployment in both high-resource and low-resource settings [1], [2], [3], [4], [5].