CGSM: Concept-Guided Segmentation Model for Precise Pulmonary Lesion Delineation

• Existing textual supervision is often coarse and location-focused, lacking boundary-relevant attributes and explicit spatial grounding for aligning descriptions with image regions.Source: §1 · pp. 1–2
• CGSM enriches coarse clinical text with LLM-generated radiology-informed concepts that provide morphological and textural priors for lesion delineation.Source: §1 · p. 2
• CVAM learns structured visual–concept correspondences to align clinical concepts with visual tokens and enforce semantic consistency.Source: §1 · p. 2
• The CM-Decoder injects aligned concepts through adaptive modulation to support morphologically grounded boundary delineation.Source: §1 · p. 2
• CGSM reports state-of-the-art performance and improves segmentation accuracy across multiple evaluation metrics.Source: §1 · p. 2

2.1 LLM-Assisted Concept Construction

• LLM-generated concept phrases capture lesion morphology, edge sharpness, texture, and spatial distribution; five clinicians rate them, and concepts averaging below 6 are discarded and regenerated.The process continues until the quality criterion is met.Source: §2.1 · p. 3

2.2 Encoder Design

• ConvNeXt-Tiny extracts four-level visual features, while frozen CXR-BERT encodes paired text and concept phrases, with a trainable projection head adapting concept representations.The same CXR-BERT backbone is used for text and concepts.Source: §2.2 · p. 3

2.3 Concept-Visual Alignment Module (CVAM)

• CVAM uses similarity-based attention to match concept tokens with visual patches, then applies bidirectional InfoNCE contrastive learning to discourage redundant grounding.The aligned concept–visual representations are provided to the decoder.Source: §2.3 · pp. 4–5

2.4 Concept Modulated Decoder (CM-Decoder)

• The CM-Decoder pools concept features to form a global descriptor that conditions visual-feature normalization and residual scaling, while multi-head cross-attention integrates text.The modulation parameters vary with the concept representation rather than remaining fixed after training.Source: §2.4 · pp. 5–6
• CGSM jointly optimizes pixel-wise cross-entropy, region-level Dice, and concept–visual alignment losses.The loss terms are balanced by weighting coefficients.Source: §2.4 · p. 6

3.1 Datasets

• QaTa-COV19 contains 9,258 annotated COVID-19 CXRs, while MosMedData+ contains 2,729 CT slices with lesion masks; both provide paired structured reports.The datasets use predefined train/validation/test splits.Source: §3.1 · p. 6

3.2 Implementation Details

• Inputs were resized to 224 × 224; the text encoder remained frozen, and experiments used AdamW with batch size 32.The learning rate followed cosine annealing from 3 × 10−4 to 1 × 10−6.Source: §3.2 · p. 6

3.3 Comparison Experiments

• 0.71% and 0.66% Dice gains over MMI-UNet on QaTa-COV19 and MosMedData+, respectively; mIoU gains are 1.21% and 0.90%.Table 1 compares methods on both datasets.Source: §3.3 · p. 7, Table 1 · p. 7
• CGSM excels in precise boundary delineation and small-lesion detection in the qualitative evaluation.Fig. 3 shows results for QaTa-COV19 and MosMedData+; green, red, and blue denote true positives, false negatives, and false positives.Source: Figure 3 · p. 7, §3.3 · p. 8

3.4 Ablation Study

• The full CGSM, combining CVAM and CM-Decoder, achieves the highest performance across ablation variants on both datasets.Adding CVAM and then CM-Decoder further improves performance over the preceding variants.Source: §3.4 · p. 8, Table 2 · p. 8

• CGSM uses radiology-informed concepts, CVAM alignment, and CM-Decoder fusion, with consistent improvements over existing methods on two public datasets.The conclusion describes concept-level priors as an effective and lightweight mechanism for multimodal medical segmentation.Source: §4 · p. 8