MCADT Model Advances Melanoma Detection

Image Credit: © DermNet

Early detection of malignant tumors, especially melanoma, is crucial for improving treatment outcomes and survival rates. Non-melanoma skin cancers, though more common, contribute less to mortality than melanoma, which remains the most lethal skin cancer.¹ Recent advancements in artificial intelligence (AI), particularly through convolutional neural networks (CNNs), have provided a significant boost in the accurate and efficient diagnosis of skin cancers, including melanoma, by analyzing medical images.^2-3

Early Detection

Observing changes in a lesion's size, shape, or color may signal the onset of melanoma. Unfortunately, diagnosis can be delayed, particularly in individuals with darker skin, resulting in poorer outcomes. In South Africa, acral lentiginous melanoma, a subtype affecting the hands and feet, has been found to be prevalent among people with darker skin tones. This highlights the importance of improving diagnostic methods for early melanoma detection across all skin types.⁴

AI in Diagnosis

AI models, particularly deep learning algorithms like CNNs, have been proven to excel in image-based diagnostics. A study using a CNN model demonstrated its ability to surpass over 130 clinicians in melanoma detection accuracy.⁵ However, while these models show promise, their "black box" nature raises concerns about interpretability, potentially leading to challenges in clinical adoption. To address this, researchers have focused on refining CNN models by employing techniques like ensemble learning, transfer learning, and data augmentation. These innovations aim to enhance model interpretability and diagnostic precision, especially in diverse datasets that include both dermoscopic and non-dermoscopic images.⁶

Study Methods and Materials

Although deep learning has advanced the field of skin cancer detection, challenges remain, such as insufficient labeled data and the overall complexity of melanoma images, often leading to inaccuracies in automated diagnosis. To mitigate these issues, recent research has focused on data augmentation, which artificially increases the size of the training dataset by applying techniques like random cropping, flipping, and blurring.⁷

The study at hand introduces a new model, the Multi-Class Augmented Deep Transfer (MCADT), which integrates multiple deep learning techniques to enhance melanoma detection. The model incorporates data augmentation methods such as RandomHorizontalFlip, RandomCrop, and GaussianBlur to address class imbalances in the dataset, particularly the underrepresentation of melanoma cases. By using a benchmark dataset, HAM10K, the model demonstrated a significant improvement in diagnostic accuracy, with results surpassing existing models and even clinical experts. The MCADT model uses a customized CNN architecture, designed to reduce overfitting and improve performance by training on augmented datasets.⁷

The model is built upon a series of convolutional layers, pooling layers, and batch normalization techniques. This design allows the model to extract features from medical images effectively. The MCADT model uses the Adam optimizer, a widely adopted optimization algorithm in deep learning, to update weights during training and improve performance. The training process uses a dataset split into training, validation, and test sets to ensure reliable evaluation of model performance. Hyperparameters, such as batch size and learning rate, were fine-tuned through experiments to achieve optimal results.

Results

Upon evaluation, the study stated the MCADT model achieved an accuracy of 93.43%, with a sensitivity of 82.05% and specificity of 88.45%. These results are on par with those of clinicians, making the model a viable tool for early melanoma detection. Additionally, the model's use of data augmentation allowed it to outperform pre-existing deep learning models, demonstrating the potential for enhanced precision in melanoma diagnosis. Researchers wrote the proposed model offers a promising approach to reducing diagnostic errors, saving time, and improving clinical outcomes for skin cancer patients.

Conclusion

The study found that AI-driven models, particularly CNNs, have the potential to revolutionize skin cancer diagnostics. By integrating advanced techniques like data augmentation and transfer learning, models like MCADT can assist clinicians in diagnosing skin lesions more accurately and efficiently. Researchers stated the success of the MCADT model in detecting melanoma, as demonstrated through testing and validation, highlights the strides being made in AI-assisted dermatology. With further improvements and validation on diverse datasets, such models could become invaluable tools in the fight against skin cancer, ultimately saving lives and reducing the burden on healthcare systems.

References

1. Patel A. Benign vs malignant tumors. JAMA Oncol. 2020;6(9):1488. doi:10.1001/jamaoncol.2020.2592
2. Saba T. Computer vision for microscopic skin cancer diagnosis using handcrafted and non-handcrafted features. Microsc Res Tech. 2021;84(6):1272-1283. doi:10.1002/jemt.23686
3. Khan MA, Sharif M, Akram T, et al. Skin lesion segmentation and multiclass classification using deep learning features and improved moth flame optimization. Diagnostics (Basel). 2021;11(5):811. doi:10.3390/diagnostics11050811
4. Kushimo OO, Salau AO, Adeleke OJ, et al. Deep learning model to improve melanoma detection in people of color. Arab j basic appl sci. 2023;30(1), 92–102. doi:10.1080/25765299.2023.2170066
5. Brinker TJ, Hekler A, Enk AH, et al. Deep neural networks are superior to dermatologists in melanoma image classification. Eur J Cancer. 2019;119:11-17. doi:10.1016/j.ejca.2019.05.023
6. Lenc K, Vedaldi A. Understanding image representations by measuring their equivariance and equivalence. Int J Comput Vis. 2019;127(5):456-476. doi:10.1007/s11263-018-1098-y
7. Sajid M, Khan AH, Malik TS, et al. Enhancing melanoma diagnostic: Harnessing the synergy of AI and CNNs for groundbreaking advances in early melanoma detection and treatment strategies. Int J Imaging Syst Technol. 2025;35: e70016. doi:10.1002/ima.70016