Diapositive 1

: Khouloud Elbedoui. Lightweight Hybrid Model Based on Compact Convolutional Transformers (CCTs) for Efficient Skin Lesions Classification.

1. Context. 2. Objectifs. 3. Methodology. 4. Results.

Context.

The Challenge: Melanoma Diagnosis. Melanoma, the most dangerous form of skin cancer, demands early and accurate automated diagnosis for improved patient prognosis..

Objectifs.

The physical examination conducted by dermatologists using the only visual diagnosis to detect anomalies or abnormal moles poses challenges for dermatologists. Achieving accurate diagnosis often necessitates expertise from highly qualified pathology specialists, yet many cases remain complex and confusing. This uncertainty can affect the reliability of results..

Introducing the Compact CNN-Transformer (CCT). We introduce an optimized hybrid architecture to address efficiency challenges in dermoscopic image diagnosis..

Methodology.

Proposed Methodology: A Robust Framework. 01. Dermoscopic Image Input.

CCT Architecture: Key Stages. The CCT model is designed to capture both local patterns and global dependencies efficiently..

CNN. Vision transformer. Efficient Net V2. Pretrained Used Architecture Models.

Balanced set of dermoscopic images of benign and malignant skin grains. Two folders with 1800 images each (224x244) of both mole types for a total of 2637 images Malignant images: 1197. Benign images: 1447. data images: 2637. The dataset is publicly available at: https://www.kaggle.com/datasets/fanconic/skin-cancer-malignant-vs-benign.

Robust data handling Is crucial for model training and unbaisd evaluation..

Hyperparameter Value Input Resolution 224×224 Embedding Dimension 256 Transformer Depth 7 Number of Heads 4 MLP Ratio 2 Batch Size 32 Optimizer AdamW Learning Rate 1e-4.

Precision. Recall. Accurarcy. F1-Score. Metrics of Validation of the classification.

Results.

The CCT model demonstrates stable convergence and robust performance in skin lesion classification..

ROC curve for the Compact CNN-Transformer Model. [image] Gemini Generated Image 2mfv1h2mfv1h2mfv.

Experimental Results : Validation & Metrics. Matrix Confusion for the Compact CNN-Transformer Model.

Precision recall F1-score support Malignant benign 80% 91% 91% 81% 85% 86% 300 360 Accuracy Macro avg Weight avgc 85% 86% 86% 85% 85% 85% 85% 660 660 660.

CCT Performance: Efficiency Meets Accuracy. The CCT achieves competitive diagnostic performance with significantly reduced computational demands..

Actual: b: benign Nalinn.OrR Probability; 96, Actual: maligurt aalignant malignant Probabiltv. 95-25% Actual Label: benign Prediction: benon Probability: 52.26% Actu.iU Label: rnafignant Predicton. malignant Probability: 63.04% Actual Label: malignant Predicton: malignant Probability: 9610% Actual Labe : martignant Prediction: benignant Probability: 95.00% Actual Label: malignant Prediction: benignar Probability 91.42* Actual Label: benign Prediction: Beniqn Probability 76.51% Actuel Label : magnant prediction: maligant probability : 94.93% Actuel Label maiigant prediction probability: 76.71%.

Comparative Analysis: CCT vs. Benchmarks. The CCT model offers an optimal balance between performance and computational cost..

Conclusion & Future Directions. The CCT model provides an optimal solution for skin lesion classification, especially in resource-constrained environments..

thanks for your attention.