ENHANCING LESION DETECTION IN CT IMAGING-A MORPHOLOGICAL CNN APPROACH WITH MULTI-IOU THRESHOLDING

One of the most helpful diagnostic and localization tools for kidney abnormalities is the computed tomography (CT) scan image. In numerous medical applications, it can give exact details regarding the location and dimensions of lesions. Traditional and manual medical testing requires a lot of labor and money. These days, the critical task of clinical diagnosis is mainly dependent on the automatic detection of abnormalities in CT scans. To advance and raise the effectiveness of medical testing, computer-aided diagnosis, or CAD, is required. The current detection algorithm's poor precision and incompleteness, however, continue to be a significant difficulty. We provide a lesion identification method based on configurational sequential convolutional networks that use a straightforward and self-explanatory measurement threshold. For the quicker region-based convolutional neural network (RCNN), we suggest integrating five intersection over union (IOU) threshold cascade RCNNs with two morphological convolution layers and a modified feature extractor that uses a single-scale image as outputs and inputs that are proportionally sized to feature maps at a stratified level to enhance the network's stability and enable better recognition of small lesions (3-5 mm). For this anomaly detection purpose, the customized convolutional net could be trained within the Pytorch framework. An average precision of 0.850 and an accuracy of 0.81 were the outcomes of the proposed methodology. The results might demonstrate that the recommended detector performed better in the dataset than Alexnet and is an excellent tool for CT lesion detection.