TL;DR
This paper introduces a novel dataset of Basal Cell Carcinoma (BCC) and Seborrheic Keratosis (SK) clinical images from the Chinese population and develops a Convolutional Neural Network (CNN) for their assisted diagnosis. The developed CNN model demonstrates performance comparable to expert dermatologists, addressing a critical need for accessible dermatological tools in regions with limited specialists.
What Problem Does This Paper Address?
Accurate diagnosis of skin lesions like Basal Cell Carcinoma (BCC) and Seborrheic Keratosis (SK) is crucial for appropriate treatment. BCC often requires surgical removal, while SK typically does not. Misdiagnosis can lead to unnecessary procedures or delayed treatment. While biopsy is the gold standard, it's invasive and often unavailable in many areas due to a shortage of pathologists. Clinically, diagnosis relies on visual appearance, medical history, and biopsy results. This paper tackles the challenge of developing a non-invasive, AI-powered diagnostic tool, specifically for the Chinese population, where existing datasets and research are scarce. The goal is to create an assistant diagnostic system that can help bridge the gap in dermatological expertise, especially in underserved regions.
Methods
The researchers employed several key techniques:
- Dataset Construction: A new dataset of BCC and SK clinical images was curated from Xiangya Hospital, Central South University. This dataset was confirmed by pathological examination, ensuring label authenticity.
- CNN Architectures: Four mainstream CNN architectures (InceptionV3, InceptionResNetV2, DenseNet121, and ResNet50) were utilized. The study compared performance when training from scratch versus fine-tuning pre-trained models (on ImageNet).
- Transfer Learning: Fine-tuning pre-trained CNN models was a core strategy to leverage existing knowledge and improve performance on the specific task of skin lesion classification.
- Model Interpretability: To understand why the CNN makes its predictions, three methods were implemented:
- t-SNE (t-distributed Stochastic Neighbor Embedding): Used to visualize high-dimensional CNN features in 2D or 3D, helping to see how well the model separates BCC and SK samples.
- Grad-CAM (Gradient-weighted Class Activation Mapping): Highlights essential regions in an input image that are important for the CNN's prediction.
- LIME (Local Interpretable Model-Agnostic Explanations): Generates saliency maps to identify important image areas, applicable to any classification model.
- Comparative Evaluation: The performance of the best-performing CNN model was directly compared against that of 21 dermatologists (8 experts, 13 generalists) on an independent test set.
Datasets
- Primary Dataset: A newly constructed Chinese skin disease dataset comprising 1456 BCC and 1843 SK clinical images. Each image is paired with corresponding medical history and pathological diagnosis. This dataset is made publicly available for academic research.
- Additional Test Set: A separate set of 116 clinical images collected from January 2019 to June 2019, used specifically for the comparative study between the CNN model and dermatologists.
Key Insights and Ideas
- Addressing a Specific Population Gap: This is the first study to focus on BCC and SK diagnosis in the Chinese population, using a dataset specifically curated for this demographic.
- Fine-tuning is Key: Fine-tuning pre-trained CNN models significantly outperformed training from scratch for this task, highlighting the power of transfer learning.
- Interpretability Matters: Beyond just accuracy, the study emphasizes understanding the CNN's decision-making process using Grad-CAM and LIME. This is crucial for building trust and potentially guiding clinical actions like biopsy site selection.
- CNNs as Diagnostic Assistants: The results suggest that CNNs can achieve diagnostic performance comparable to experienced dermatologists, offering a valuable tool for augmenting clinical practice, especially in resource-limited settings.
Key Results
- Best Performing Model: The fine-tuned InceptionResNetV2 achieved the highest performance, with an accuracy of 0.855 and an Area Under the Curve (AUC) of 0.919 on the test set.
- CNN vs. Dermatologists: The InceptionResNetV2 model achieved an AUC of 0.937, outperforming the average performance of general dermatologists and demonstrating comparability to expert dermatologists. Specifically, the model's optimal TPR and FPR were 0.897 and 0.103, respectively, which was comparable to the average of expert dermatologists.
- Interpretability Success: Grad-CAM and LIME visualizations generally located lesion areas and provided visual clues, suggesting their utility in understanding model predictions and potentially guiding biopsy site selection.
Limitations and Future Work
- Dataset Size: The authors acknowledge that the quantity of clinical images, while substantial, is not massive. More images are needed to verify the model's generalization performance.
- Multimodal Data: The current model only uses clinical images. Future work will incorporate patient metadata (gender, age, lesion location, medical history) to create a more comprehensive, multimodal diagnostic system.
- Clinical Trials for Applications: The proposed applications, particularly recommending biopsy locations using LIME, require clinical trials to verify their feasibility and effectiveness in real-world scenarios.
- Generalization: Further efforts will focus on collecting more diverse samples to improve the model's generalization capabilities.
Summary updated