Smart Diagnosis and Early Intervention in PCOS: A Deep Learning Approach to Women's Reproductive Health

Polycystic Ovary Syndrome (PCOS) is a widespread disorder in women of reproductive age, characterized by a hormonal imbalance, irregular periods, and multiple ovarian cysts. Infertility, metabolic syndrome, and cardiovascular risks are long-term complications that make early detection essential. In this paper, we design a powerful framework based on transfer learning utilizing DenseNet201 and ResNet50 for classifying ovarian ultrasound images. The model was trained on an online dataset containing 3856 ultrasound images of cyst-infected and non-infected patients. Each ultrasound frame was resized to 224x224 pixels and encoded with precise pathological indicators. The MixUp and CutMix augmentation strategies were used to improve generalization, yielding a peak validation accuracy of 99.80% by Densenet201 and a validation loss of 0.617 with alpha values of 0.25 and 0.4, respectively. We evaluated the model’s interpretability using leading Explainable AI (XAI) approaches such as SHAP, Grad-CAM, and LIME, reasoning with and presenting explicit visual reasons for the model’s behaviors, therefore increasing the model’s transparency. This study proposes an automated system for medical picture diagnosis that may be used effectively and confidently in clinical practice.

💡 Research Summary

The paper presents a deep‑learning framework for early detection of Polycystic Ovary Syndrome (PCOS) using ovarian ultrasound images. The authors collected a publicly available dataset of 3,856 ultrasound frames, equally divided between PCOS‑positive (infected) and PCOS‑negative (non‑infected) cases. Each image was resized to 224 × 224 pixels, normalized, and subjected to rigorous data cleaning (duplicate removal, outlier handling, categorical encoding, scaling). To improve generalization, the authors applied two modern augmentation techniques: MixUp (α = 0.25) and CutMix (α = 0.4), which blend images and labels in a controlled manner, thereby expanding the effective training distribution and reducing overfitting.

Two transfer‑learning architectures were evaluated: DenseNet201 and ResNet50, both initialized with ImageNet weights. DenseNet201’s dense connectivity promotes feature reuse and mitigates vanishing gradients, while ResNet50’s residual blocks enable stable training of deeper networks. Both models retain standard convolution‑batch‑norm‑ReLU blocks, dropout, and a final sigmoid‑activated dense layer for binary classification. Training employed the Adam optimizer (learning rate = 1e‑4), binary cross‑entropy loss, batch size = 32, and early stopping (patience = 15 for DenseNet, 10 for ResNet). DenseNet was trained for 98 epochs, ResNet for 67 epochs, with checkpointing to preserve the best weights.

Performance results are striking. DenseNet201 achieved a validation accuracy of 99.80 % and a validation loss of 0.617, ultimately reaching 100 % test accuracy (test loss = 4.54 × 10⁻⁵). The confusion matrix shows perfect classification of all 162 PCOS‑positive and 223 PCOS‑negative cases, yielding precision, recall, and F1‑score of 1.00 for both classes. ResNet50 attained a test accuracy of 97.02 % (test loss = 2.11 × 10⁻⁶) with 154/162 PCOS‑positive cases correctly identified and zero false positives; eight false negatives indicate a small residual error, possibly due to overlapping visual features or label noise. Training and validation curves for both models remain tightly coupled, suggesting minimal overfitting.

To address the “black‑box” nature of deep networks, the authors incorporated Explainable AI (XAI) methods: SHAP, Grad‑CAM, and LIME. Grad‑CAM heatmaps consistently highlighted cystic regions within the ovary, providing visual confirmation that the models focus on clinically relevant structures. SHAP values quantified pixel‑level contributions, while LIME offered local surrogate explanations for individual predictions. These visual and quantitative explanations enhance clinician trust and facilitate model validation in a medical context.

The paper also discusses limitations. The dataset originates from a single public source, limiting external validity; class imbalance statistics are not fully disclosed; and the study lacks an assessment of inference speed, hardware requirements, or integration into clinical workflows. Future work is suggested to include multi‑center data collection, multimodal fusion of ultrasound with hormonal and demographic data, model compression for edge deployment, and prospective clinical trials to evaluate real‑world impact.

In summary, the study demonstrates that transfer‑learning with DenseNet201 (and, to a slightly lesser extent, ResNet50) combined with MixUp/CutMix augmentation can achieve near‑perfect classification of PCOS from ultrasound images. The addition of SHAP, Grad‑CAM, and LIME provides transparent, clinically interpretable explanations, positioning the system as a promising low‑cost, scalable screening tool for resource‑constrained settings where expert radiologists are scarce.