Enhanced Classifications of Skin  Melanoma, Actinic Keratosis, and Basal Cell Carcinoma Tumors Versus Normal Tissues Utilizing  YOLOv11 and  EfficientNetV2M  Deep Learning Models

Technical Summary: Enhanced Classifications of Skin Melanoma, Actinic Keratosis, and Basal Cell Carcinoma Tumors Versus Normal Tissues Utilizing YOLOv11 and EfficientNetV2M Deep Learning Models

Problem Statement
Skin tumors represent a significant global health challenge, with invasive melanoma accounting for the majority of skin cancer-related deaths despite comprising only a small fraction of total cases. Current automated diagnostic systems utilizing dermoscopic data face three persistent obstacles: severe class imbalance in public datasets (biasing models toward benign cases), high visual similarity between benign and malignant lesions complicating feature discrimination, and a lack of consistent cross-architecture benchmarking. Furthermore, while recent detection-oriented architectures (like the YOLO family) and classification-oriented models (like EfficientNet) have advanced computer vision, their specific applicability and comparative performance in binary skin lesion classification—particularly under clinical class-imbalance conditions—have not been systematically characterized.

Methodology
The study utilizes the HAM10000 dataset, a public collection of 10,015 dermatoscopic images covering seven distinct skin lesion classes. To reflect clinical reality, the authors deliberately preserved a 4:1 class imbalance between benign and malignant cases. The dataset was split into training (80%), validation (10%), and test (10%) sets, ensuring no patient data leakage across partitions.

The research evaluates three distinct deep learning configurations:

1. YOLOv11: A fine-tuned YOLOv11s model (Ultralytics) adapted for binary classification. The model underwent full fine-tuning using an SGD optimizer, a learning rate of 0.001, and a batch size of 32 over 50 epochs. Data augmentation included ±45° rotation, 50% horizontal/vertical flipping, and 50% scaling.
2. EfficientNetV2M + Neural Network (NN): The EfficientNetV2M backbone was frozen as a feature extractor (pre-trained on ImageNet), feeding into a custom neural network head consisting of five dense layers with batch normalization and dropout. This configuration was trained using an SGD optimizer.
3. EfficientNetV2M + Support Vector Machine (SVM): The same EfficientNetV2M backbone was used to extract features, which were then standardized and classified using a polynomial SVM (degree 1) from the scikit-learn library.

The study conducted ablation studies on optimizers (AdamW vs. SGD), learning rates, batch sizes, and augmentation policies to identify the most influential design choices.

Key Contributions

• YOLOv11 Benchmarking: This work presents the first benchmark of YOLOv11 for binary skin lesion classification, achieving high performance without architectural modifications or knowledge distillation.
• Hybrid Pipeline Evaluation: The authors evaluate and compare hybrid pipelines pairing EfficientNetV2M with both neural network and SVM classifiers, analyzing the trade-offs between deep learning end-to-end training and traditional machine learning classification on extracted features.
• Rigorous Comparative Analysis: The study provides a reproducible baseline against state-of-the-art methods (including ResNet50, RegNetY-320, and DSNet) using consistent dataset splits and evaluation metrics.
• Ablation Insights: Systematic identification of hyperparameters that most significantly influence classification performance under class-imbalanced conditions.

Results

• YOLOv11: Achieved the highest test accuracy of 92.8%, with a recall of 86.5% and specificity of 93.3%. The model demonstrated a strong capacity for minimizing false negatives, making it suitable for high-sensitivity screening.
• EfficientNetV2M + NN: Achieved a test accuracy of 85.1%. While it showed effective generalization on the minor class, it yielded lower recall and F1 scores compared to the SVM variant.
• EfficientNetV2M + SVM: Achieved a test accuracy of 85.0% with an F1 score of 83.0%. This configuration offered the highest specificity (96.8%) and precision (83.8%), providing a more conservative precision-recall profile suitable for high-precision triage.
• Comparative Performance: YOLOv11 outperformed several established baselines, including ResNet50 (90%), RegNetY-320 (91%), and DSNet (91.7%). The EfficientNet hybrid approaches performed comparably to other CNNs like MobileNet and InceptionV3 found in prior literature.

Significance and Claims
The paper claims that the choice of architecture should be guided by the specific operational clinical context rather than a single "best" model.

• YOLOv11 is positioned as the superior candidate for high-sensitivity screening where minimizing missed malignancies (false negatives) is paramount.
• EfficientNetV2M + SVM is presented as a robust alternative for high-precision triage, offering superior specificity to reduce unnecessary follow-up procedures.

The authors conclude that while YOLOv11 establishes a new state-of-the-art for overall accuracy in this domain, the hybrid EfficientNet approaches offer flexible, modular alternatives that balance precision and recall effectively. The study emphasizes that these results provide reproducible benchmarks for future research, though they acknowledge the need for prospective evaluation on independent clinical datasets to validate generalizability beyond the ISIC benchmark.