Background The meniscus is a crucial structure in the knee joint, and its abnormalities are best detected using MRI. However, manual interpretation of MRI scans is time-consuming and prone to inter-observer variability. With advancements in artificial intelligence (AI), deep learning models offer a promising approach to automate evaluation, improving diagnostic accuracy and efficiency. Objective This study evaluates the performance of a deep learning-based Mask R-CNN model for the segmentation and classification of the medial meniscus in MRI scans. Unlike prior studies that used bounding box-based segmentation of knee structures, our model utilizes precise polygonal annotations to ensure pixel-wise segmentation limited to the meniscus, allowing accurate abnormality detection. Methods We used a dataset of 3,600 sagittal proton density-weighted fat-suppressed (PD-FS) MRI images. The meniscus was manually annotated using the VGG Image Annotator (VIA) tool, focusing on accurate delineation of the meniscus while excluding adjacent anatomical structures. The Mask R-CNN model, with a ResNet-50 backbone and feature pyramid network (FPN), was trained for 50 epochs using a structured dataset. We evaluated performance using metrics such as area under the curve (AUC), segmentation accuracy, sensitivity, and specificity. Results The model demonstrated high performance in distinguishing normal and abnormal menisci. It achieved an AUC of 0.992 for normal menisci and 0.963 for abnormal menisci. Segmentation results confirmed precise meniscus delineation, validating the exclusion of non-meniscal regions and improving classification accuracy. Training and validation loss trends showed effective learning without overfitting, supporting the model's generalization capability. Conclusion The Mask R-CNN model provides an accurate AI-assisted tool for segmenting and classifying MRI scans. By using pixel-wise segmentation rather than bounding boxes, our approach minimizes the inclusion of surrounding structures, ensuring more refined and clinically relevant abnormality detection. Focusing solely on the meniscus enables targeted abnormality detection while reducing the radiologists' workload. Future work will focus on multi-center dataset validation and expanding the model to sub-classify meniscal abnormalities for enhanced clinical applicability.