Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Toward Auditable Neuro-Symbolic Reasoning in Pathology: SQL as an Explicit Trace of Evidence

Kewen Cao, Jianxu Chen, Yongbing Zhang, Ye Zhang, Hongxiao Wang

TL;DR

The paper tackles the opacity of AI in pathology by introducing an auditable, SQL-centered reasoning framework that converts visual features into structured, executable queries. It decouples reasoning from end-to-end classification, with Global/Local Feature Reasoning Agents generating SQL, a Knowledge Comparison Agent validating against diagnostic knowledge, and a Report Agent fusing results with CNN evidence. Experiments on mu-bench and GADVR show the approach yields competitive accuracy and significantly improved interpretability via auditable SQL traces linking cellular measurements to diagnoses. This work demonstrates the practicality of SQL as a symbolic interface for trustworthy AI in pathology and suggests paths for integrating external medical knowledge.

Abstract

Automated pathology image analysis is central to clinical diagnosis, but clinicians still ask which slide features drive a model's decision and why. Vision-language models can produce natural language explanations, but these are often correlational and lack verifiable evidence. In this paper, we introduce an SQL-centered agentic framework that enables both feature measurement and reasoning to be auditable. Specifically, after extracting human-interpretable cellular features, Feature Reasoning Agents compose and execute SQL queries over feature tables to aggregate visual evidence into quantitative findings. A Knowledge Comparison Agent then evaluates these findings against established pathological knowledge, mirroring how pathologists justify diagnoses from measurable observations. Extensive experiments evaluated on two pathology visual question answering datasets demonstrate our method improves interpretability and decision traceability while producing executable SQL traces that link cellular measurements to diagnostic conclusions.

Toward Auditable Neuro-Symbolic Reasoning in Pathology: SQL as an Explicit Trace of Evidence

TL;DR

The paper tackles the opacity of AI in pathology by introducing an auditable, SQL-centered reasoning framework that converts visual features into structured, executable queries. It decouples reasoning from end-to-end classification, with Global/Local Feature Reasoning Agents generating SQL, a Knowledge Comparison Agent validating against diagnostic knowledge, and a Report Agent fusing results with CNN evidence. Experiments on mu-bench and GADVR show the approach yields competitive accuracy and significantly improved interpretability via auditable SQL traces linking cellular measurements to diagnoses. This work demonstrates the practicality of SQL as a symbolic interface for trustworthy AI in pathology and suggests paths for integrating external medical knowledge.

Abstract

Automated pathology image analysis is central to clinical diagnosis, but clinicians still ask which slide features drive a model's decision and why. Vision-language models can produce natural language explanations, but these are often correlational and lack verifiable evidence. In this paper, we introduce an SQL-centered agentic framework that enables both feature measurement and reasoning to be auditable. Specifically, after extracting human-interpretable cellular features, Feature Reasoning Agents compose and execute SQL queries over feature tables to aggregate visual evidence into quantitative findings. A Knowledge Comparison Agent then evaluates these findings against established pathological knowledge, mirroring how pathologists justify diagnoses from measurable observations. Extensive experiments evaluated on two pathology visual question answering datasets demonstrate our method improves interpretability and decision traceability while producing executable SQL traces that link cellular measurements to diagnostic conclusions.
Paper Structure (11 sections, 3 figures, 2 tables)

This paper contains 11 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Multi-scale Feature Database Construction
  4. The SQL-centered Feature Reasoning Branch
  5. Knowledge-based Hypothesis Validation
  6. CNN Branch and Report Generation
  7. Experiments
  8. Datasets and Experiment Settings
  9. Comparison with VLM Models
  10. Ablation Studies
  11. Conclusions

Figures (3)

  • Figure 1: Overview of the proposed framework. The framework couples a SQL-reasoning branch built on a multi-scale feature database with a CNN model for visual reasoning. This SQL branch employs Feature Reasoning Agents to formulate auditable queries and a Knowledge Comparison Agent to validate the results against diagnostic criteria into a hypothesis, which is then fused with the CNN output by a Report Agent for a final diagnostic prediction with an auditable reasoning chain.
  • Figure 2: Schema of the Multi-Scale Database. Features are grouped into Local (left) and Global (right) levels, forming the explicit evidence base for SQL reasoning.
  • Figure 3: Example of SQL-based reasoning correcting a CNN error.Top left: H&E patch from GADVR. Top right: segmentation overlay highlighting neoplastic cells (red).