Early detection of subtle pathological features in medical images is critical for improving patient outcomes but remains challenging due to low contrast, small lesion size, and limited annotated data. The research contribution is a hybrid attention-enhanced CNN specifically tailored for small object detection across mammography, CT, and retinal fundus images. Our method integrates a ResNet-50 backbone with a modified Feature Pyramid Network, dilated convolutions for contextual scale expansion, and combined channel–spatial attention modules to preserve and amplify fine-grained features. We evaluate the model on public benchmarks (DDSM, LUNA16, IDRiD) using standardized preprocessing, extensive augmentation, and cross-validated training. Results show consistent gains in detection and localization: ECNN achieves an F1-score of 88.2% (95% CI: 87.4–89.0), mAP@0.5 of 86.8%, IoU of 78.6%, and a low false positives per image (FPPI = 0.12) versus baseline detectors. Ablation studies confirm the individual contributions of dilated convolutions, attention modules, and multi-scale fusion.However, these gains involve higher computational costs (≈2× training time and increased memory footprint), and limited dataset diversity suggests caution regarding generalizability. In conclusion, the proposed ECNN advances small-object sensitivity for early disease screening while highlighting the need for broader clinical validation and interpretability tools before deployment.