Technical Summary: Singpath-VL

1. Problem Statement

The application of Multimodal Large Language Models (MLLMs) in cytopathology, specifically cervical cytology screening based on The Bethesda System (TBS), has been significantly hindered by a critical bottleneck: the scarcity of large-scale, high-quality, publicly available datasets that pair cytology images with detailed, expert-level morphological descriptions.

Unlike histopathology, which relies on tissue architecture, cervical cytology diagnosis depends on the precise interpretation of subtle cellular and nuclear features (e.g., nuclear hyperchromasia, irregular nuclear membranes, increased nuclear-to-cytoplasmic ratio) from individual cells. General-purpose MLLMs lack the fine-grained visual grounding required for this task, and existing manual datasets are too small to train specialized models effectively.

2. Methodology

The authors propose Singpath-VL, a specialized vision-language assistant for cervical cytology. The methodology consists of two core components: a novel data generation pipeline and a specialized multi-stage training strategy.

A. Data Generation Pipeline (Singpath-CytoText)

To overcome data scarcity, the team developed a three-stage automated pipeline to synthesize a million-scale, high-fidelity image-description dataset from an internal repository of cervical cytology image tiles:

1. Parallel Image Caption Generation: Three distinct, powerful open-source MLLMs (Qwen3-VL-32B, InternVL-3.5-38B, and Baichuan-Omni-1.5) are queried in parallel to generate preliminary descriptions based on TBS-relevant morphological dimensions.
2. Consensus Fusion and Coherent Summarization: A Large Language Model (LLM) acts as an integrator to:
   • Extract consensus features described by the majority of models.
   • Filter out "model-missing dimensions" (features consistently misidentified due to model limitations).
   • Synthesize a single, fluent, and structured paragraph.
3. Expert Knowledge Injection: A smaller, domain-specific expert model (fine-tuned on cervical cytopathology) analyzes the original image and the fused description to supplement critical, fine-grained details overlooked by generalist models, ensuring final accuracy.

B. Model Training Strategy

The base model selected is Qwen3-VL-4B. It is fine-tuned using a three-stage recipe to balance domain adaptation with the retention of general knowledge:

1. Vision-Language Alignment: Full-parameter fine-tuning on the generated Singpath-CytoText dataset to establish grounding between visual features (cell morphology, staining) and pathological terminology.
2. Supervised Fine-Tuning (Instruction Following): Reformating data into instruction-following formats using structured TBS observation templates, simulating real-world diagnostic queries and multi-turn dialogues.
3. Knowledge Replay: To mitigate catastrophic forgetting, the model is trained on a mix of:
   • Domain-specific data (cytopathology QA pairs).
   • General-domain data (sampled from the original Qwen3-VL capabilities).
   • Rationale: Using the base model for data generation in this stage acts as regularization to preserve the model's original reasoning framework.

3. Key Contributions

1. Novel Data Pipeline: Introduction of an automated, three-stage consensus-and-refinement pipeline that leverages multiple MLLMs as weak annotators and an expert model for refinement, successfully generating a million-scale dataset (Singpath-CytoText).
2. Specialized Model (Singpath-VL): Development of a cytopathology-specific MLLM via a multi-stage training regimen that effectively balances deep domain adaptation with general knowledge retention.
3. Open-Source Contribution: Commitment to open-sourcing a portion of the synthetic dataset and benchmarks to advance research in computational cytopathology.

4. Experimental Results

The model was evaluated on two custom benchmarks: MorphoPercept-Bench (fine-grained morphological feature detection) and CytoCell-Bench (TBS diagnostic classification).

Morphological Perception (MorphoPercept-Bench)

• Performance: Singpath-VL achieved an average accuracy of 89.0%, significantly outperforming all baselines (Qwen3-VL-4B: 53.1%, InternVL-3.5-38B: 50.0%).
• Expert Comparison: The model's accuracy exceeded the inter-rater agreement level of human experts (78.2%), demonstrating superior consistency in identifying features like Nuclear Enlargement (91.2%) and Nuclear Atypia (95.7%).

Diagnostic Classification (CytoCell-Bench)

• Performance: Singpath-VL achieved an average accuracy of 80.2% across six TBS categories (NILM, ASC-US, LSIL, ASC-H, HSIL, AGC).
• Comparison:
  • It matched the performance of the specialized EfficientNet B0 baseline on the NILM category (100.0%).
  • It outperformed EfficientNet B0 in clinically ambiguous categories: +5.5% on ASC-US and +3.7% on ASC-H.
  • General-purpose MLLMs failed catastrophically (e.g., Qwen3-VL-4B achieved 16.7% average accuracy), often showing a bias toward specific classes.

Qualitative Analysis

Visual comparisons confirmed that Singpath-VL provided morphologically accurate descriptions and correct TBS classifications for both Negative (NILM) and Low-Grade (LSIL) cases. In contrast, general models frequently misidentified morphological features, leading to incorrect diagnoses and a tendency to over-classify samples as positive.

5. Significance and Future Directions

Significance:
Singpath-VL represents a paradigm shift in computational cytopathology by demonstrating that specialized MLLMs can surpass both general-purpose models and human inter-rater consistency in fine-grained morphological tasks. It addresses the "data scarcity" bottleneck through a scalable, synthetic data generation approach, proving that high-fidelity domain-specific AI assistants can be built without massive manual annotation.

Limitations & Future Work:

• Scope: Currently limited to cervical cytology; generalization to other subspecialties (e.g., thyroid, urine) requires retraining.
• Tasks: Validation is focused on description and classification; future work aims to integrate clinical history and slide-level screening.
• Reasoning: Future iterations aim to move from pattern recognition to causal reasoning, where the model explicitly articulates the logical chain of inference (feature identification $\to$ rule application $\to$ diagnosis) to enhance transparency and auditability.
• Expansion: Plans to expand into a "Pan-Cytopathology" assistant and develop interactive, human-in-the-loop diagnostic support systems.