Sumit Shekhar Reviewer
Approved
Relevance and Originality
The manuscript addresses an important healthcare problem by proposing an automated framework for hepatitis prediction using ultrasound images. The integration of lightweight deep learning with adaptive texture encoding reflects an effort to balance computational efficiency and diagnostic accuracy. The topic is highly relevant within medical image analysis and artificial intelligence in healthcare. The originality is moderately strong, particularly in combining APRN U, granular texture encoding, and Bayesian neural networks, although similar hybrid approaches have been explored in recent studies.
Methodology
The study presents a well structured pipeline, beginning with preprocessing, followed by dual branch feature extraction, feature fusion, dimensionality reduction, and classification. The inclusion of APRN U and CLAHE for preprocessing is appropriate for ultrasound images. The hybrid feature extraction using CNN and texture descriptors is methodologically sound. However, certain implementation details, such as dataset size, annotation process, and hyperparameter tuning strategy, require clearer elaboration to ensure reproducibility.
Validity and Reliability
The results demonstrate strong performance with reported accuracy and AUC values. The inclusion of ablation study enhances confidence in the contribution of each component. Nevertheless, the dataset appears limited and lacks discussion on diversity and external validation. The absence of cross dataset evaluation raises concerns about generalizability in real clinical environments.
Clarity and Structure
The manuscript follows a logical organization, with clearly defined sections and progression of ideas. Figures and tables support the explanation effectively. However, there are several grammatical inconsistencies and formatting issues, including uneven spacing and sentence construction, which should be corrected. Some sections are overly descriptive and could be more concise.
Results and Analysis
The performance comparison with existing methods is useful and demonstrates the effectiveness of the proposed model. The discussion highlights improvements across multiple metrics. However, deeper critical analysis of why the model outperforms others would strengthen the paper. Including statistical validation or confidence intervals would further improve the robustness of the findings.
Sumit Shekhar Reviewer