An Integrated Deep Learning Model to Analyze CT Scans for Minimally Invasive Accurate Classification of T1a Small Renal Masses

Convolution and attention-based deep learning techniques are emerging as promising approaches for automated medical image classification and semantic segmentation for minimally invasive diagnosis of malignant masses in vital organs. However, current deep learning based automated classification techniques are suitable only for detecting renal masses greater than 4.0 cm in diameter, which have a higher probability of malignancy increasing the morbidity, costly disease-management, and the probability of early mortality. T1a Small renal masses (SRM) are less than 4.0 cm in diameter and have a much smaller probability of being malignant and are corrected with partial nephrectomy or ablation. Detection of SRMs poses significant challenges due to low contrast with the surrounding and irregular shapes. As the imaging techniques improve, automated medical image-analysis techniques need further improvement for diagnosing SRMs. In this study, we present an integrated deep learning model and an algorithm for diagnosing T1a SRMs. The model integrates convolution blocks with residual links for capturing local features without information loss, cross-channel attention to focus on important features, principal component analysis to prune low variance features for reduced error and improved computational efficiency, and a transformer-encoder comprising self-attention to maintain dependencies and global context across image-patches. We trained and evaluated our model on the KiTS19 and KiTS21 challenge datasets for three ranges {1.2 – 2.0, 2.0 – 3.0, 3.0 – 4.0}. The model achieved an accuracy of 98.1%, 98.5%, and 97.3% for the three ranges, respectively. The accuracy of classification results significantly outperforms other approaches.