Abstract
Reflectance confocal microscopy (RCM) is an effective non-invasive tool for cancer diagnosis. However, acquiring and reading RCM images requires extensive training and experience, and novice clinicians exhibit high discordance in diagnostic accuracy. Quantitative tools to standardize image acquisition could reduce both required training and diagnostic variability. To perform diagnostic analysis, clinicians collect a set of RCM mosaics (RCM images concatenated in a raster fashion to extend the field view) at 4-5 specific layers in skin, all localized in the junction between the epidermal and dermal layers (dermal-epidermal junction, DEJ), necessitating locating that junction before mosaic acquisition. In this study, we automate DEJ localization using deep recurrent convolutional neural networks to delineate skin strata in stacks of RCM images collected at consecutive depths. Testing our model against an expert labeled dataset of 504 RCM stacks, we achieved 88.07% classification accuracy and a nine-fold reduction in the number of anatomically impossible errors compared to the previous state-of-the-art.
What Problem Does This Paper Address?
Reflectance Confocal Microscopy (RCM) is a powerful, non-invasive "optical biopsy" tool used to diagnose skin cancer by allowing clinicians to see individual cells in real-time. The most critical area for diagnosis is the Dermal-Epidermal Junction (DEJ)—the boundary where approximately 80% of skin cancers originate.
Currently, finding the DEJ is a manual "pilot" step. A clinician must scroll through a 3D stack of images (an RCM stack) to identify the depth of the epidermis, the DEJ, and the dermis before they can capture high-resolution wide-field "mosaics" for analysis. This process is highly subjective, requires extensive training, and leads to significant variability between operators. To make RCM a standard clinical tool, a method to automate this "strata delineation" accurately and consistently is required.
Methods
The authors propose a hybrid architecture called a Recurrent Convolutional Network (RCN) with a specialized attention mechanism. The pipeline works in two main stages:
- Spatial Feature Extraction: A modified Inception-V3 CNN extracts textural features from individual 2D RCM slices. Since RCM images are grayscale and rely on refractive index differences, the CNN learns to recognize the specific "texture" of different skin layers.
- Sequential Modeling: Because skin layers follow a strict biological order (Surface $\rightarrow$ Epidermis $\rightarrow$ DEJ $\rightarrow$ Dermis), the authors used Bidirectional Gated Recurrent Units (GRUs). This allows the model to consider the context of images above and below a specific slice to make a decision.
- Toeplitz Attention: This is the paper's primary algorithmic contribution. Standard "Global Attention" looks at the entire image stack to classify one slice, which can be noisy. The authors introduced Toeplitz Attention, which constrains the model to look only at a local neighborhood of slices using a learnable, depth-independent weight kernel. This mimics how a human expert focuses on local transitions to find a boundary.
Datasets
The study utilized a large and diverse dataset for this task:
- 504 RCM stacks containing a total of 21,412 expert-labeled images.
- Data was collected from five different institutes.
- The dataset included both benign and suspicious lesions (e.g., melanomas and nevi). This is vital because cancerous tissue often disrupts normal skin architecture, making it harder for simpler models to classify.
Key Insights and Ideas
- The Anatomical Constraint: Classification accuracy is not the only metric that matters. A model might be accurate but predict a "dermis" slice between two "epidermis" slices—a biological impossibility. By using RCNs, the authors enforced the monotonicity of skin layers.
- Local vs. Global Context: While global attention helps eliminate impossible transitions, the Toeplitz (local) attention was superior for precisely localizing the DEJ-Dermis boundary. It turns out that for skin, "where you are" relative to your immediate neighbors is more informative than the total stack depth.
Key Results
- Accuracy: The best model (Toeplitz Attention with $D=1$) achieved 88.07% classification accuracy.
- Error Reduction: Compared to previous state-of-the-art methods, the RCN approach achieved a nine-fold reduction in anatomically impossible errors.
- Boundary Precision: The Mean Absolute Error (MAE) for identifying the Epidermis-DEJ boundary was only 5.76 $\mu m$, which is roughly the thickness of a single cell layer.
Limitations and Future Work
- Interpretability: While the RCN performs well, it is less "interpretable" than traditional regression models. In a medical setting, clinicians often want to know why a model chose a specific boundary.
- Optical Aberrations: The model struggles more with the deeper DEJ-Dermis boundary because light scatters more at depth, leading to "fuzzier" images. Improving resolution at depth remains a hardware challenge.
- Real-time Integration: The next step is to integrate this model into RCM hardware to guide clinicians during live imaging sessions, potentially reducing inter-operator variability and standardizing imaging procedures.
Summary updated