Prostate Cancer Detection using Large Vision Models on Apparent Diffusion Coefficient MRI

Prostate cancer is among the most common cancers in men across the world, and early detection is considered crucial for enhanced health outcomes. This approach performs a computer-aided diagnosis of prostate cancer specifically on MRI-derived apparent diffusion coefficient (ADC) maps. Two proposed approaches are evaluated: the first uses feature extraction with Convolutional Neural Networks (CNNs), including GoogleNet, DenseNet-121, and DarkNet-19, and uses traditional Machine Learning Algorithms like SVM (including various kernels), Logistic Regression, and Random Forest for classification. The second approach leverages transfer learning on pre-trained CNNs, including GoogleNet and DenseNet121. Various metrics were evaluated to identify the best performance, including accuracy, precision, recall, $\mathbf{F 1}$ score, and ROC AUC. The results demonstrate that transfer learning and CNN-based feature extraction are indeed capable of achieving good results, which also points to a potentiality for deep learning in supporting clinical diagnostics. This comparative analysis demonstrates the feasibility of machine learning as a useful tool in the detection of various forms of cancers, the integration of which into radiological workflows can greatly enhance the reliability and quickness of diagnoses.