Multi-Modal Cancer Detection Systems: Advances in Early Diagnosis
Abstract — Early and accurate cancer diagnosis is critical for improving treatment outcomes and reducing the mortality rate. This study presents a deep learning-based multimodal cancer detection framework designed for the analysis of ultrasound, MRI, and CT scan images of the breast, liver, and thyroid organs. The proposed system utilizes EfficientNet-based architectures for the binary classification of cancerous and noncancerous medical images. EfficientNetB0 was employed for smaller ultrasound datasets to achieve computational efficiency and reduce overfitting, whereas EfficientNetB3 was utilized for more complex MRI and CT scan datasets to enhance feature extraction capability and classification performance. To improve model generalization, EfficientNetB3 was pretrained on approximately 117,000 medical images across 135 classes before fine-tuning on the target datasets. The framework was implemented using Python and PyTorch, along with supporting libraries, including OpenCV, NumPy, Pillow, and Torchvision. The experimental evaluation demonstrated promising results, achieving accuracies of 94% for breast ultrasound, 90% for liver ultrasound, 85% for thyroid ultrasound, 85% for breast MRI, 93% for liver MRI, and 80% for liver CT scan images. The results indicate that the proposed multimodal approach combined with transfer learning can effectively improve medical image classification and support early cancer diagnosis. Keywords — Multi-modal Cancer Detection, Deep Learning, EfficientNet, Transfer Learning, Medical Image Analysis, Ultrasound, Magnetic Resonance Imaging (MRI), Computed Tomography (CT).