RECNN-CGAN: A Resource-Efficient Deep Learning Framework for Multi-Class Mammogram Classification and Diagnosis

A resource-efficient convolutional neural network with conditional generative adversarial networks (RECNN-CGAN) is proposed for mammogram classification and diagnosis (MCD) to achieve accurate and computationally efficient breast cancer diagnosis.

The system employs Gaussian filtering for noise removal, while the Non-Subsampled Contourlet Transform (NSCT) extracts discriminative features from mammograms at multiple scales, preserving directional information. Handcrafted mammographic features are fused with deep representations learned by the proposed dual-branch RECNN, which incorporates a lightweight pooling-integrator mechanism to reduce model complexity while retaining discriminative capability. This design addresses the computational complexity associated with conventional CNNbased mammography classifiers. To address class imbalance and reduce overfitting, the CGAN module generates effective representations of abnormal mammograms. The CGAN module was incorporated to enhance the robustness and generalization of our model. Experiments were performed on three heterogeneous datasets: DDSM (2,620 images), MIAS (325 images), and a private clinical dataset collected from Bharat Scans, Chennai, India, between July 2022 and January 2023 (488 images), encompassing variations in patient characteristics and imaging conditions.

The proposed model achieved per-class sensitivities of 99.11%, 99.61%, and 99.05%, and specificities of 99.09%, 99.55%, and 99.00% for normal, benign, and cancerous classes, respectively. The accuracy values for the normal, benign, and cancerous cases were 99.06%, 99.47%, and98.95%, respectively. The macro-averaged sensitivity, specificity, accuracy, precision, and F1-score were 99.57%, 99.32%, 99.41%, 98.72%, and 98.75%, respectively. The segmentation component achieved an IoU of 0.982, indicating strong agreement with boundary-level expert annotations.

Dataset-wise per-class evaluation on independent non-augmented test subsets revealed detection rates for the cancerous class to be consistently high across all datasets, thereby minimizing false negatives and false positives, which are critical for early diagnosis and clinical reliability in breast cancer detection.

Overall, the RECNN-CGAN framework demonstrated superior performance compared with the selected comparative methods, including multitask GCN, ResNet50 + InceptionV3, and CNN-based feature selection approaches, while maintaining lower computational requirements. With diagnostic accuracy, resource efficiency, and robustness across datasets, the proposed system retains a strong chance of integration into large-scale breast cancer screening workflows, particularly when implemented in low-resource or high-throughput clinical settings.