Lung Tumor Segmentation in Medical Imaging Using U-NET

Tumors are a deadly condition often triggered by a range of abnormal modifications and genetic abnormalities. Early tumor diagnosis is essential due to the highly concerned nature of the disease. Early detection and treatment of tumors can significantly reduce mortality rates. This paper presents a model for tumor segmentation in medical imaging that uses the U-NET architecture to increase precision. The model’s encoding and decoding processes have been applied with skip connections to boost performance while simplifying model training. Images were cropped around the lower abdominal regions, and all images used in the study were then resized to 256*256 pixels for standardization. The proposed model deals with the class imbalance using data augmentation and oversampling. The experiments achieved a dice score of 0.853±0.02; F-score of 0.905±0.02; and a sensitivity of 0.897±0.02, compared with various existing models. As part of the model’s application, the pytorch-lightning library is used to successfully identify lung cancer scans, thereby proving to be a precise and efficient method of tumor identification. Accordingly, the study emphasizes the accuracy and speed of the applied model as a useful instrument for the earliest detection of tumors. The proposed approach helps to achieve more relevant and accurate segmentation and thus provides enhancements in medical images analysis if such challenges as an imbalance data set are well handled.