A Fusion Based Convolutional Neural Network method for Renal Clear Cell Carcinoma (ccRCC) detection and localization from multiphase Computed Tomography (CTs) and a Database for classification and segmentation of ccRCC cases

The aim of this thesis was the creation of a curated database of abdominal CT scans, including both normal and pathological renal cases, and the development of a Convolutional Neural Network (CNN), with a U-Net core for detection, segmentation, and multi-class classification of renal lesions. This retrospective study was conducted using data from G. Papanikolaou Hospital of Thessaloniki database, following approval from the scientific committee of the hospital. All data were fully anonymized and manually annotated, resulting in 390 cases: 90 histologically confirmed clear-cell renal cell carcinoma (ccRCC) cases, 150 Bosniak type I renal cysts, and 150 normal cases without focal lesions. To address computational constraints, a single representative frame from each CT study was extracted in JPEG format along with its corresponding mask to form the training dataset. Based on previous works, a U-net model was appropriately adjusted for kidney detection and segmentation, whereas a CNN was developed to classify cases into three categories. Despite using whole-abdomen frame inputs and the predominance of small lesions, the algorithm achieved satisfactory segmentation and classification performance, with 0.35 Dice Similarity Coefficient for segmentation and 56% classification accuracy. An additional script processed the entire CT series frame by frame in order to simulate a real scenario for possible clinical application. This enabled the detection of focal lesions and distinction between RCC and cysts with satisfying results. Interestingly, the model achieved higher sensitivity in detecting RCC cases, which is more clinically relevant than misclassifying normal results. This study demonstrates the potential of CNN-based renal lesion detection and classification using real-world data, providing a structured dataset that can contribute to further research. Although challenges remain, particularly with small lesion detection and multiphase fusion, the results are promising for further studies in AI-assisted renal imaging.