1.1 Research contributions

The main results of this investigation are summarized as follows.1.This research highlights the image enhancement and systematic data augmentation strategy's contribution to the stabilization of the models and the improvement of classification performance, particularly on heterogeneous histopathological datasets.2.This research also conducts controlled experiments to measure the effect of preprocessing techniques and data augmentation strategies and proves them to be critical to mitigating data scarcity and enhancing generalization.3.The study thoroughly compares the most recent CNNs; VGG19, Inception, InceptionResNet, and NasMobileNet—concluding why lightweight NAS-based models (e.g., NasMobileNet) outperform deeper.4.This work also puts forward an interpretable preprocessing framework to facilitate histopathology image enhancement to support osteosarcoma diagnosis and demonstrate its potential to reduce pathologists' workload, facilitate increased consistency in diagnosis, and offer scalable AI-supported solutions to mixed healthcare settings.

This paper's remaining sections are arranged as follows. In Section 2, pertinent literature is reviewed with an emphasis on the state of deep learning in medical imaging today. The study's methodology, dataset, preprocessing procedures, and augmentation strategies are all covered in Section 3. The experimental data and assessment metrics are presented in Sections 4 and 5, while a thorough explanation of all the experimental outcomes is given in Section 5. With conclusions and suggestions for potential future study topics, Section 6 concludes the study.

3.1 Dataset collection

The Osteosarcoma Tumor Assessment dataset provided by UT Southwestern^36–38 UT Dallas was used in this research. This dataset was specifically curated for viable vs. necrotic tumor assessment, making it highly relevant for diagnostic applications. Although the dataset was derived from only four patients, each whole-slide image was divided into numerous non-overlapping tiles capturing diverse tissue morphologies. This tiling approach enhances intra-patient variability. Table 2 describes the dataset with its classes, annotations and split sizes. Fig. 2 demonstrates some of the original images collected from the data source.

Close

Fig. 2 Original Images from the dataset.

Export to PPT

This dataset reflects the true heterogeneity of osteosarcoma histology, where the same patient may exhibit viable, necrotic, and non-tumor tissue. Such variability makes automated diagnosis challenging but also clinically essential.

3.2 Image preprocessing

Histopathology images often suffer from uneven staining, noise artifacts, poor contrast, and variability in illumination—factors that significantly degrade deep learning (DL) model performance. Before augmentation and training, we implemented a multi-stage preprocessing pipeline to enhance image quality while preserving diagnostic features.

Noise reduction was obtained by an amalgamation of median, Gaussian, and bilateral filtering. Such operations effectively reduced salt–and–pepper noise, smoothened high–frequency artifacts, and retained structural edges like tumor boundaries and stromal textures. This guaranteed that learning by downstream models occurred on histological features and not on acquisition noise. Histogram equalization and Contrast–Limited Adaptive Histogram Equalization (CLAHE) were used to overcome poor contrast and uneven illumination, respectively, and improved tissue differentiation and nuclear visibility. Sharpening with unsharp masking further highlighted morphological boundaries important for pathology interpretation. Segmentation refinement by Otsu's adaptive thresholding and follow–up by morphological operations (erosion, dilation, opening, closing) to remove staining debris and improve tumor–non–tumor separation finally occurred. All the images were also resized to 224 × 224 pixels to conform to CNN input requirements and normalized to the range [0,1] intensity range to obtain stable convergence on the network. Post–processing visualization are presented in Fig. 3 which demonstrate better nuclei definition and sharper tumor edges. Owing to presentation requirements to be concise, mathematical expressions for filtering and enhancement operations have been shifted to the Supplementary Section and are available for reference.

Close

Fig. 3 Enhanced Images from dataset.

Export to PPT

Histopathological images usually consist of challenges such as uneven staining, noise artifacts, suboptimal contrast, and lighting variations (attributes that remarkably compromise the accuracy of DL models). Before performing augmentation and learning, we created a multi-stage preprocessing workflow aimed at image quality improvement and key diagnostic property preservation.

3.3 Image augmentation

For enhanced generalization under limited data, systematic data augmentation was used. This comprised geometric transformations (rotation, flipping, shifting), imagemetric changes (intensity scaling), and elastic distortions to inject controlled diversity. Although these transformations boost resistance to common acquisition and staining variations, they do not perfectly simulate the underlying biological diversity of osteosarcoma tissue. The intention behind the controlled data augmentation plan was to work out the incremental contribution of synthetic diversity to model stability and not to recapitulate the full biological heterogeneity of the disease. Once enhanced images are received onto the system, image augmentation is conducted. Due to the limited size and lack of balance of the dataset, image augmentation was implemented only on the training set to ensure the addition of heterogeneity and the prevention of overfitting. Augmentation injects synthetic diversity while retaining key histopathological characteristics. We implemented four controlled settings depicted in Fig. 4. To assess the effect of dataset expansion on model generalization, three controlled augmentation scales were designed—650, 1000, and 1500 images per class. These levels were selected based on preliminary experiments that indicated smaller increments produced minimal improvement, while larger expansions led to feature redundancy and overfitting. This stratified design enables a balanced evaluation of model performance under progressively enriched data diversity. The scenario I includes basic rotation of ±30° along with horzontal/vertical fliping for 2543 images. In scenario II, 3744 images wee used with batic rotation same as scenario I along with shifting of (±10°) with scaling of 0.8–1.2 times. The scenario III includes settings of scenario III with elastic deformation and intesity scaling for 5,9636 images. For the rest of 860 Images, baseline scenario IV was considered which does not have any augmentation.

Close

Fig. 4 Image augmentation scenarios.

Export to PPT

Mathematically, a transformed image I′ is obtained by applying the transformation T using Eq (1):Eq (1)I′(x,y)=I(T(x,y))where T( includes rotation Rθ, translation (tx,ty)), scaling s, and flipping F.

Overfitting is a major challenge in medical imaging tasks, especially when datasets are limited, as models may memorize training data rather than learning generalizable features. To address this, multiple regularization strategies were incorporated, including dropout, batch normalization, L2 regularization, and early stopping. To further improve generalization, overfitting mitigation was performed using dropout, batch normalization, L2 regularization, and early stopping, using Eqs (2)–(4).

Randomly deactivate neurons with probability p.Eq (2)y=11−p·m⊙f(x)where m∼Bernoulli(1−p)Eq (3)xˆ=x−μBσB2+ϵEq (4)Ltotal=LCE+λ∑wi2

Augmentation size settings (650, 1000, and 1500) were chosen following pilot experiments to represent increasingly growing but computationally manageable expansions. Such levels captured enough diversity to examine the trend to generalize without causing too much synthetic duplication to prejudice the learning process. The augmented images are presented in Fig. 5, which demonstrates the changes to the original image after enhancemnet and augmentation.

Close

Fig. 5 Augmented images.

Export to PPT

3.4 Transfer learning models

In this step, the four pre-trained CNNs(InceptionV3, InceptionResNetV2, VGG19, and NasMobileNet) are fine tunned. These architectures were selected for their complementary characteristics. VGG19 was used as a deep but simple architecture for baseline performance. Whereas InceptionV3 and ResNet have Multi-scale feature extraction, better gradient flow. The NasMobileNet is NAS-optimized, lightweight, efficient, and suited for limited-resource clinical deployment. Each model's final dense layer was replaced with a softmax classifier using Eq (5):Eq (5)P(y=k∣x)=ezk∑j−1Kezk,K=4where K = 4 (Non-tumor, Viable, Non-viable-Tumor, viable:non-viable).

Training was performed using Adam optimizer (η=10−4) batch size 32, and categorical cross-entropy loss using Eq (6):Eq (6)LCE=−∑i=1N∑k=1Kyiklogyikˆ

To systematically evaluate the influence of dataset size on classification performance, we designed multiple augmented training batches derived from the preprocessed dataset. Rather than applying augmentation uniformly, we generated controlled subsets to examine how incremental increases in training volume impacted generalization ability. Hyperparameters were determined through a restricted empirical search balancing model stability and overfitting control. Batch sizes between 2 and 64 were tested to accommodate GPU memory limits and dataset size, while the number of epochs (50) was chosen based on early convergence observations during pilot runs. This setup ensured consistent optimization conditions across all architectures. The hyperparameters, including batch size (2–64) and number of epochs (50), were optimized through a constrained grid search designed to balance performance stability and computational efficiency. The selected ranges were guided by prior histopathology studies, which indicate that excessively large batch sizes may reduce generalization on small datasets. The grid search was intentionally limited to avoid overfitting and to maintain fair comparison across models with differing architecture complexities. The first configuration served as a baseline, where the raw preprocessed dataset of 860 images was used without augmentation, providing a benchmark against which subsequent scenarios could be compared. A second configuration moderately expanded the dataset with an additional 650 synthetic samples (approximately 1510 images in total), created using simple transformations such as rotations and flips to introduce limited variability. The third configuration further increased the dataset size to roughly 1860 images by adding 1000 augmented samples that included both geometric transformations and mild intensity variations, exposing the models to structural and staining differences frequently encountered in histopathology slides. Finally, the fourth configuration represented the most extensive augmentation, expanding the dataset to about 2360 images with an additional 1500 synthetic samples. This scenario integrated elastic deformations and stronger intensity scaling, simulating biologically realistic distortions and inter-laboratory variability. Collectively, these systematically designed scenarios provided a structured framework to investigate how augmentation scale influences both robustness and diagnostic accuracy of deep learning models in osteosarcoma classification. The progressive augmentation sizes allowed us to examine the trade-off between dataset volume and model stability. Smaller expansions (650) primarily tested whether even minimal augmentation can reduce overfitting. Intermediate expansions (1000) represented a balance between computational cost and robustness. Larger expansions (1500) examined whether excessive synthetic diversity further improved generalization or risked introducing artifacts. By structuring augmentation in this stepwise manner, we could directly quantify the impact of dataset size on osteosarcoma classification and isolate whether performance gains were due to image enhancement, dataset expansion, or model architecture.

4.1 Experimental setup

Using TensorFlow Keras, all experiments were conducted on a Linux workstation equipped with an NVIDIA RTX 3060 GPU (8 GB RAM). While splitting the dataset, a balanced class distribution was preserved with the proportions of 75 % for training, 12 % for validation, and 13 % for testing. For transfer learning, We considered the VGG19, InceptionV3, InceptionResNetV2, and NasMobileNet models which were pre-trained on ImageNet. To facilitate a fair comparison across the models, a hyperparameter adjustment was conducted for each model, as illustrated in Table 3.

4.2 Evaluation parameters

TensorFlow Keras. For the dataset, I used a 75 %-training, 12 %-validation, and 13 %-testing split that preserves the class distribution. For transfer learning, the ImageNet pre-trained models including VGG19, InceptionV3, InceptionResNetV2, and NasMobileNet, were assessed. Hyperparameters were tuned for each of the models to facilitate fair comparisons, as shown in Table 3. Traditional evaluation parameters were used to assess the performance of the proposed framework.

Accuracy (ACC): Accuracy (ACC) is defined by the proportion of correctly classified instances compared to all predictions made, as shown in Eq (13). Even with the frequent reporting of accuracy, it can provide an incorrect impression if the dataset is heavily imbalanced, so other metrics are critical.Eq (13)ACC=TP+TNTP+TN+FP+FN

Sensitivity (Recall/True Positive Rate): Measures the model's ability to correctly classified positive instances, Eq (14). Undetected tumors represent the most critical false negatives. Hence, a high sensitivity value is essential clinically.Eq (14)SE=TPTP+FN

Specificity (SP): Provides an assessment of correctly classifying negative instances (i.e., non-tumor tissue) in Eq (15). High specificity is crucial for eliminating false positive results and thereby avoiding unnecessary procedures.Eq (15)SP=TNTN+FP

Precision (Positive Predictive Value): Is proof of the positivity prediction made and reduction of overdiagnosis probability, Eq (16).Eq (16)Precision=TPTP+FP

F1-Score: Creates a balance between recall and precision harmonically, Eq (17). This is crucial for the dataset that has imbalanced classes.Eq (17)F1=2·(Precision·Recall)Precision+Recall

Along with the Confusion Matrices, which explain the performance metrics by the various classes (Non-tumor, Viable, Necrosis) and helped to identify areas of misclassification and the impacts of class imbalance, we also used ROC to illustrate the sensitivity and specificity relationship. Secondly, ROCs illustrated the trade-off between sensitivity and specificity across varying thresholds. The ROC quantified overall discriminative ability, with values closer to 1.0 indicating superior classification reliability. Thirdly, ablation study and statitical anlysis using t-test and MCC helped to understand the alidaity of the obtained results. MCC is calculated using Eq (18).Eq (18)MCC=(TP×TN)−(FP××FN)(TP+FP)(TP+FN)(TN+FP)(TN+FN)

4.3 Ablation study: augmentation scenarios

The four controlled augmentation scenarios were tested to investigate the influence of dataset expansion: Scenario I (650 samples/class), Scenario II (1000 samples/class), Scenario III (1500 samples/class), and Scenario IV (no augmentation). Each of the scenarios was applied to the same preprocessed dataset to enable the isolation of the augmentation effects. Each model was assessed through confusion matrices and ROC curves to gain insights into their classification capabilities and to identify areas of improvement. The different augmented scenarios and their respective number of images are illustrated in Fig. 6.

Close

Fig. 6 No. of Images across four scenarios.

Export to PPT

4.3.1

4.3.1 Scenario I (650 samples/class)

In Scenario I, each class was expanded to approximately 650 samples using basic augmentation techniques such as flipping and rotation. Results indicated that all models achieved accuracies above 92 %, with NasMobileNet and InceptionV3 reaching 92.96 % as explained through comparison in Table 4. Fig. 7 (a, b, c, d) presnts confusion matrices revealed frequent misclassification between viable and necrotic tissues. ROC values also confirmed high discriminative performance, particularly for NasMobileNet (0.96), as demonstaretd in Fig. 8 (a, b, c, d).

Table 4 Evaluation metrics for scenario I.

Model	Accuracy	Specificity	Sensitivity	F1 Score
NasMobileNet	92.96 %	93 %	93 %	92 %
VGG19	92.25 %	91 %	92 %	91 %
InceptionResNetV2	92.25 %	91 %	92 %	91 %
InceptionV3	92.96 %	92 %	93 %	92 %

Close

Fig. 7 (a) NasMobileNet, (b) VGG19, (c) Inception ResNetV2, (d) InceptionV3 Confusion matrix for secnario 1.

Export to PPT

Close

Fig. 8 (a) NasMobileNet, (b) VGG19, (c) Inception ResNetV2, (d) InceptionV3 ROC for scenario I.

Export to PPT

4.3.2

4.3.2 Scenario II (1000 samples/class)

Scenario II expanded the dataset to 1000 samples per class, incorporating geometric transformations and mild intensity scaling. This yielded the best overall performance, with NasMobileNet achieving 95.07 % accuracy, sensitivity, and specificity as shown in Table 5. Both Inception-based models achieved ∼93.6 %, while VGG19 remained slightly lower at 92.9 %. Fig. 9(a and b,c,d) and Fig. 10 (a, b, c, d) demonstrate the Confusion matrices and ROC curves for all trained models, respectively. The moderate augmentation is observed and which improves model generalization most effectively.

Table 5 Evaluation metrics for scenario II.

Model	Accuracy	Specificity	Sensitivity	F1 score
NasMobileNet	95.07 %	95 %	95 %	95 %
VGG19	92.96 %	93 %	93 %	93 %
InceptionResNetV2	93.66 %	94 %	94 %	94 %
InceptionV3	93.66 %	94 %	94 %	94 %

Close

Fig. 9 Confusion matrix for scenario II (a) NasMobileNet, (b) VGG19, (c) Inception ResNetV2, (d) InceptionV 3.

Export to PPT

Close

Fig. 10 ROC for scenario II (a) NasMoIbileNet, (b) VGG19, (c) Inception ResNetV2, (d) InceptionV3 4.3.3 Scenario III (1500 samples/class).

Export to PPT

In Scenario III, augmentation was increased to 1500 samples per class, including elastic deformations and stronger intensity scaling. While InceptionResNetV2 and InceptionV3 performed best here (94.37 %), NasMobileNet's accuracy plateaued at 92.96 %, and specificity dropped slightly, as shown in Table 6. This suggests that excessive augmentation can introduce artifacts that reduce robustness for lightweight models like NasMobileNet. Fig. 11 (a, b, c, d) demonstrates a confusion matrix, and Fig. 12(a and b,c,d) presents ROC for scenario III.

Table 6 Evaluation metrics for scenario III.

Model	Accuracy	Specificity	Sensitivity	F1 Score
NasMobileNet	92.96 %	89 %	93 %	92 %
VGG19	90.14 %	87 %	90 %	89 %
InceptionResNetV2	94.37 %	95 %	94 %	94 %
InceptionV3	94.37 %	95 %	94 %	94 %

Close

Fig. 11 NasMoIbileNet, (b) VGG19, (c) Inception ResNetV2, (d) InceptionV3 Confusion matrix for scenario III.

Export to PPT

Close

Fig. 12 ROC curves for (a) NasMobileNet, (b) VGG19, (c) Inception ResNetV2, (d) InceptionV3 for scenario III.

Export to PPT

4.3.3

4.3.3 Scenario IV (No augmentation)

In the baseline condition without augmentation, performance was noticeably weaker across all models. NasMobileNet achieved only 89.44 % accuracy, while VGG19, InceptionResNetV2, and InceptionV3 ranged from 90 to 92 % as shown in Table 7. The reduced scores highlight the critical role of augmentation in mitigating data scarcity and preventing overfitting in medical imaging tasks. Confusion matrix and ROC curve for scenario 4 are presented in Fig. 13 (a, b, c, d) and Fig. 14 (a, b, c, d), respectively.

Table 7 Evaluation metrics for scenario IV.

Model	Accuracy	Specificity	Sensitivity	F1- Score
NasMobileNet	89.44 %	89 %	89 %	89 %
VGG19	90.85 %	87 %	91 %	90 %
InceptionResNetV2	91.55 %	87 %	92 %	90 %
InceptionV3	92.96 %	93 %	93 %	93 %

Close

Fig. 13 NasMoIbileNet, (b) VGG19, (c) Inception ResNetV2, (d) InceptionV3 Confusion matrix for scenario IV.

Export to PPT

Close

Fig. 14 (a) NasMoIbileNet, (b) VGG19, (c) Inception ResNetV2, (d) InceptionV3 ROC for scenario IV.

Export to PPT

4.4 Statistical analysis

In order to check if model performance differences were due to random chance, the four architectures, VGG19, InceptionV3, InceptionResNetV2, and NasMobileNet, described in all augmentation scenarios, were fully analyzed. Model performance results from the five different cross-validation folds for each architecture were analyzed using paired t-tests. The bootstrapping method was used to produce 95 % CIs using 1000 iterations. To measure the quality of each model's classification and to take into account the effect of class imbalances on classification, the Matthews Correlation Coefficient (MCC) was also calculated. The results are explained in Table 8, which shows that MCC is higher for NasMobileNet, moderate for Inception-based models; however, it is slightly lower for VGG19. This represented an index of balanced classification quality for this research.

Most of the comparisons were statistically significant (i.e. p-values <0.05) multiple times, which is shown in Table 9, implying that, at the 95 % confidence level, the performance differences observed across models were not significant. All p-values >0.05 indicate no statistically significant difference between model performances. Although NasMobileNet exhibited the highest mean accuracy (95.07 % ± 0.34 %) in Scenario II, its advantage over InceptionV3 (93.63 % ± 0.41 %) and InceptionResNetV2 (94.37 % ± 0.39 %) did not reach statistical significance (p > 0.05). For the VGG19 model, the mean accuracy was very slightly lower (92.90 % ± 0.46 %), still overlapping the confidence intervals of the other networks, thus reiterating the statistical equivalence of the results. The MCC values ranging from 0.88 to 0.92 continued to support the consistent prediction reliability across the different architectures. The absence of statistical significance must be because of the small dataset size (with only four patient sources), which intrinsically limits inter-sample variability, thus, drives the power of inferential tests down. In small-sample biomedical imaging studies, high intra-class correlation and overlapping tissue morphologies often lead to non-significant p-values, even when mean accuracy differences appear numerically distinct. Thus, while the results demonstrate methodological stability and reproducibility, they do not yet justify claims of superiority for any single architecture. This emphasizes the need for larger, multi-institutional datasets to achieve statistically robust differentiation between deep-learning models in future work.

Although NasMobileNet consistently reached slightly higher accuracy and specificity, these improvements were not statistically significant. Thus, these models appear to have comparable diagnostic ability given the augmentation and preprocessing used. This may indicate that, when considering preprocessing and augmentation, the depth of the network or number of parameters may not be the primary underlying driver of performance. The confidence intervals support this conclusion as demonstated in Table 10. NasMobileNet has the smallest 95 % CI (0.9502–0.9534), indicating little variance and a high degree of reliability over the runs. In contrast, InceptionResNetV2 has a wider CI (0.6434–0.9452), which indicates variability and possible instability resulting from the larger parameter size and sensitivity to sparse data.

These results overall validate that there is no consistent material difference in performance between the models. Despite NasMobileNet having the greatest mean accuracy, the statistically weak dataset and limited diversity in patients contribute more to this uncertainty in the results, and consequently, to generalization. Therefore, the accuracy figures should indicate consistent methodologies rather than a definitive clinical superiority.

4.5 Comparative summary across scenarios

To consolidate findings, Table 11 presents the best performance achieved by each model across all scenarios. NasMobileNet excelled in Scenario II (1000 samples/class), while Inception-based models peaked under Scenario III (1500 samples/class). The results also suggest that lightweight models prefer moderate augmentation, whereas deeper models benefit from larger, more diverse datasets. The comparative analysis is done in Fig. 12 which clearly demonstrates the higher performance among models with different scenarios. Table 12 presents the performance of models demonstarting across differenet scenarios.).

4.6 Comparison with baseline methods

Finally, to place our results in context, we compared them with previous works in osteosarcoma classification. While prior studies (e.g., Anisuzzaman et al. Vezakis et al.) reported accuracies between 91 and 95 %, they lacked systematic ablation studies or robust preprocessing strategies. Our approach, combining preprocessing, augmentation, and transfer learning, achieved superior robustness and interpretability as shown in Table 13. Fig. 15 presents comparison of precision of pre-trained models used across different scenarios. The performance of TL models used in all scenarios are presented in Fig. 16.

Close

Fig. 15 Precision for pre-trained TL models used for each scenario.

Export to PPT

Close

Fig. 16 Precision of TL used for each scenario.

Export to PPT

5.1 Significance and clinical perspective

The present study has various takeaways that can benefit both computational pathology and translational AI design. First, the study includes a controlled ablation platform that assesses the relative value of preprocessing and augmentation; an evaluation seldom performed in explicit osteosarcoma imaging. Second, the in-study measures of performance (up to 95 % sensitivity/specificity) suggest potential clinical practicality for screening or triage applications, in which false negatives are to be minimized. Third, the exhibited effectiveness of NAS-based lightweight architectures advocates for deployment in settings where there is no advanced computational infrastructure, thereby increasing the availability of AI-aided diagnostics. The current study is not an attempt at introducing a new deep learning architecture; it is, however, an orderly investigation of established CNN architectures under constrained preprocessing and augmentation settings. The innovation is the methodological rigidity of distinguishing the individual contribution of each component of the pipelines, as well as how lightweight architectures, like NasMobileNet, can reproduce similar discriminative accuracy at lower computational complexities. The annotations for the database were reviewed by two board-certified physicians under a consensus protocol. However, the inter-observer variability was numerically not evaluated, while it can induce labeling bias. Furthermore, the limitation of the absence of external validation by means of multi-institutional data entails future efforts to incorporate cross-institutional databases and official agreement metrics in an attempt to increase clinical reliability. These results are grounded in an available public set consisting of four cases. Extensive tiling was performed in an attempt to capture intra-tumoral heterogeneity, yet the low patient diversity restricts generalizability. In the future, validation with much larger, independent, and multicenter sets shall be performed to establish robustness.

5.2 Limitations and future work

This study, while providing a structured and systematic evaluation of preprocessing, augmentation, and transfer learning strategies for osteosarcoma histopathology classification, is subject to several limitations that must be acknowledged. The primary limitation is the restricted dataset size, comprising images derived from only four patients in a single-center source. Although thousands of non-overlapping tiles were generated to capture intra-tumoral variability, such a small cohort inherently limits model generalizability and increases the risk of overfitting. The current train-validation-test split may introduce data leakage, as patches from the same patient could appear across subsets, potentially inflating reported performance metrics. Additionally, the study's statistical power remains limited, with non-significant p-values (all p > 0.05) across models indicating that apparent performance differences likely arise from data constraints rather than true architectural superiority. While data augmentation helped simulate morphological diversity, excessive transformations occasionally produced non-biological tissue distortions, emphasizing the need for biologically constrained augmentation in histopathological pipelines. The reliance on standard architectures (VGG19, InceptionV3, InceptionResNetV2, and NasMobileNet) and conventional augmentation techniques, though methodologically sound, limits the degree of algorithmic innovation. Furthermore, the absence of external dataset validation and the lack of quantified inter-observer variability among pathologists restricts clinical translation. Future research should focus on expanding to multi-institutional datasets to capture broader biological variability and assess.