<![CDATA[Cervical cancer remains one of the leading causes of cancer-related mortality among women worldwide, particularly in low- and middle-income countries where access to trained specialists and advanced diagnostic infrastructure is severely limited. Cervical Intraepithelial Neoplasia (CIN), the precancerous lesion preceding cervical malignancy, exists across a spectrum of severity — CIN1, CIN2, and CIN3 — each requiring distinct clinical management. Accurate and timely differentiation of these stages through colposcopy is critical; however, the subjective nature of visual interpretation, high inter-observer variability, and the scarcity of experienced colposcopists in resource-constrained settings frequently lead to misclassification, delayed treatment, or unnecessary interventions. These diagnostic limitations underscore an urgent need for automated, objective, and reliable computer-aided diagnosis (CAD) tools capable of supporting clinicians in multiclass cervical lesion grading. To address this challenge, this study proposes a hybrid CAD framework for multiclass cervical lesion classification using colposcopy images. The proposed method integrates handcrafted texture and color features — specifically Color Moments and Gray Level Run Length Matrix (GLRLM) — with deep semantic features extracted from pretrained EfficientNetB3 and MobileNetV3 networks. Classification is performed by a fully connected Deep Neural Network (DNN) operating on the aggregated feature representations to categorize images into four classes: Normal, CIN1, CIN2, and CIN3. Experiments were conducted on an IARC colposcopy dataset comprising 200 cases, employing stratified 5-fold cross-validation to ensure robust evaluation. The proposed methodology achieved an average accuracy of 82.67% ± 3.65% and a Cohen's Kappa of 0.7466 ± 0.0634, with a peak accuracy of 86.22% on the best-performing fold. Results demonstrate strong recognition performance for Normal and CIN3 classes, while intermediate stages remain more challenging due to overlapping visual characteristics. Overall, the hybrid feature fusion strategy enhances classification robustness and shows promise for supporting reliable cervical precancer screening in resource-limited healthcare environments.]]>
