Cardiac tumors present significant challenges in terms of early diagnosis and treatment planning due to their low incidence and complex anatomical location. Precise segmentation of CT scan images plays an essential role in enhancing the efficacy of diagnosis and clinical decision support. This paper introduces a deep learning-oriented automatic segmentation pipeline for identifying and outlining cardiac tumors with high accuracy. The proposed pipeline starts with the preprocessing of the CT scans, such as intensity normalization, denoising, resampling, and cropping, to make the images uniform and receptive to better visualize the tumors. A CNN encoder–decoder architecture motivated from U-Net is adopted to extract multi scale spatial and contextual cues for robust and dense tumor segmentation, and skip connections maintain structural cues during the process of reconstruction. To achieve optimal learning and compensate for class imbalance, a hybrid loss function integrating Binary Cross-Entropy, Soft Dice Loss, and Focal Tversky Loss is utilized. Furthermore, the traditional machine learning classifiers Naïve Bayes and K-Nearest Neighbors are used in addition to the outputs of CNN for complementary classification and generalizability improvement. The pipeline was trained and tested on a CT-based image dataset for 70:10:20 partitioning, and experimental results revealed robust performance, obtaining an accuracy of 97.8%, precision of 95.71%, recall of 93.62%, F1-score of 94.63%, Dice coefficient of 94.52%, IoU of 91.83%, and AUC of 96.45%. Overlay of segmentation revealed precise delimitation of tumoral outlines, and ROC curve analysis further confirmed model robustness, too. These results proved the efficacy of the proposed CNN encoder–decoder pipeline with hybrid loss optimization and complementary classifiers, qualifying it as a reliable and clinically feasible method for automatic segmentation of cardiac tumors in CT scans.