Study purpose: Li and Shen (2018, Sensors) propose two specialized deep learning networks - the Lesion Indexing Network (LIN) for segmentation and the Lesion Feature Network (LFN) for dermoscopic feature extraction and classification - developed for the three tasks of the ISIC 2017 Skin Lesion Analysis Challenge.
LIN architecture: The Lesion Indexing Network combines two Fully Convolutional Residual Networks (FCRN-88) in a novel two-stage pipeline followed by a Lesion Index Calculation Unit that converts network predictions into a structured binary segmentation mask.
LFN architecture: The Lesion Feature Network uses a patch-based CNN approach that extracts local texture and color features from skin lesion patches, combining spatial feature maps with global average pooling for simultaneous dermoscopic feature localization and melanoma classification.
Key results: LIN achieved Jaccard index 0.753 for segmentation and AUC 0.912 for classification; LFN achieved AUC 0.848 for dermoscopic feature extraction, representing competitive performance in the ISIC 2017 challenge leaderboard across all three tasks.
FCRN-88 backbone: The network is built on FCRN-88, a fully convolutional residual network with 88 layers that captures multi-scale spatial features through a sequence of convolutional, pooling, and residual skip connection blocks, inheriting ResNet's gradient stability during deep network training.
Two-stage pipeline: LIN operates in two stages: the first FCRN-88 module generates a coarse segmentation map, which is then fed as additional input to the second FCRN-88 module for refinement, enabling iterative boundary sharpening without requiring manual post-processing.
Lesion Index Calculation Unit: A specialized post-processing module converts the continuous probability maps from the second network stage into a binary segmentation mask, applying threshold optimization to maximize Jaccard index rather than using a fixed 0.5 cutoff.
Training strategy: The network was trained on the ISIC 2017 training set of 2000 dermoscopy images with expert-annotated segmentation masks, using data augmentation including random rotation, flipping, and color jitter to improve generalization to test images.
Patch-based approach: Rather than processing whole dermoscopy images at once, the LFN extracts fixed-size patches from segmented lesion regions, allowing the network to focus on local texture and color patterns within the lesion without confusion from surrounding normal skin.
Multi-scale feature extraction: Patches are extracted at multiple spatial scales to capture both fine-grained dermoscopic texture features (pigment network details, dot patterns) and broader structural features (asymmetry, border irregularity) relevant to melanoma classification.
Shared architecture for feature detection and classification: The LFN uses a common feature extraction backbone for both dermoscopic feature detection (Task 2: globules, streaks, pigment network) and melanoma classification (Task 3), exploiting the semantic relationship between dermoscopic features and diagnostic categories.
Feature map aggregation: Spatial feature maps from the final convolutional layer are globally average-pooled into a fixed-length feature vector, which feeds into separate task-specific output heads for simultaneous feature detection and classification prediction.
Task 1 segmentation (LIN): LIN achieved a mean Jaccard index of 0.753 on the ISIC 2017 segmentation test set, competitive with the challenge top-10 performance (approximately 0.762 for the challenge winner), demonstrating that the two-stage FCRN approach is an effective segmentation strategy.
Task 3 classification (LIN): Applying lesion segmentation masks from LIN as preprocessing improved downstream classification, with the classification component achieving AUC 0.912 for melanoma vs. seborrheic keratosis vs. nevus three-class discrimination.
Task 2 feature detection (LFN): The LFN achieved AUC 0.848 averaged across dermoscopic feature detection tasks, with variable performance across feature types - globules and pigment network detection performed best while milia-like cysts and streaks remained more challenging.
Combined pipeline performance: Using LIN for segmentation followed by LFN for classification in a sequential pipeline modestly improved classification AUC compared to using LFN alone, confirming the utility of segmentation as a preprocessing step for feature-based classification.
Pipeline architecture value: The LIN-LFN approach demonstrates that decomposing the diagnostic problem into segmentation, feature detection, and classification subtasks - mirroring how dermatologists analyze dermoscopy images - outperforms end-to-end black-box classification alone.
Feature detection as interpretability: By explicitly predicting dermoscopic features (globules, streaks, pigment network), LFN provides more interpretable AI decisions than pure classification networks, as dermatologists can understand which features drove the melanoma prediction.
Segmentation prerequisite: The improvement in classification when using LIN-generated segmentation masks suggests that automatic lesion isolation is a clinically important preprocessing step, supporting the case for deploying segmentation AI as part of any clinical dermoscopy decision support system.
Multi-task efficiency: Sharing a feature extraction backbone for both dermoscopic feature detection and melanoma classification enables model deployment with a single forward pass, an important consideration for mobile or low-latency clinical deployment.
Two-stage pipeline complexity: Running LIN followed by LFN increases inference time compared to single-network approaches; for real-time clinical deployment on mobile devices, architectural efficiency improvements or network distillation would be needed.
Patch extraction dependency on segmentation quality: LFN performance depends on LIN segmentation quality - segmentation errors propagate into incorrect patch extraction regions, potentially degrading feature detection and classification. Segmentation failure modes on challenging lesions (low contrast, hair artifacts) warrant further analysis.
ISIC 2017 dataset limitations: The three-class ISIC 2017 dataset (melanoma, nevus, seborrheic keratosis) does not represent the full spectrum of clinically relevant skin lesion diagnoses; models trained on this dataset may not generalize to other diagnostic categories.
Future directions: Attention-based architectures that jointly learn which regions to segment and which features to extract in a single unified model, combined with clinical metadata incorporation, represent natural extensions of this work toward real-world clinical deployment.