Fine-grained Medical Diagnosis Beyond Binary Classification

This direction is merged into another direction: diagnostic granularity. Its last version stays readable and citable here.

Version 1 of 1162 papers167 conceptsfrozen 2026-10-04

In clinical practice, a diagnosis of "abnormal" is rarely enough to determine a treatment plan. While early medical AI focused on binary classification—distinguishing healthy tissue from diseased—modern research has shifted toward fine-grained classification. This direction seeks to differentiate between visually similar disease subtypes, etiologies, and severity levels to provide actionable clinical insights.

The Limitations of Binary Screening

Most automated diagnostic systems act as "abnormality flags." However, as noted in Next-Gen Pneumonia Detection: A Review of Secure AI, CNN Models, and Federated Learning, identifying the presence of pneumonia is only the first step. The critical clinical question is whether the infection is bacterial or viral, as this determines whether antibiotics are appropriate.

Kapal Dev and colleagues argue that grouping diverse pathogens into broad categories masks the specific clinical features needed for effective triage. Similarly, in oncology, Zhongyi Han points out that simple "benign vs. malignant" labels fail to capture the eight specific subtypes of breast cancer that dictate different surgical and therapeutic schedules.

Hierarchical and Multi-Scale Architectures

To achieve this granularity, researchers are moving away from monolithic classifiers toward hierarchical and multi-scale models. The intuition is that broad features identify the disease family, while specialized "expert" modules or high-resolution branches identify the subtype.

  • Divide and Conquer: Kapal Dev proposed a Hierarchical Convolutional Network (HCN) that uses a "divide and conquer" strategy to separate COVID-19 from other viral pneumonias.
  • Multi-Scale Integration: Bum-Chae Kim developed the MGI-CNN, which uses parallel streams to "zoom in" and "zoom out" on lung nodules. This allows the model to distinguish between solid and ground-glass nodules—a distinction vital for predicting malignancy.
  • Unified Hierarchies: Models like CancerUniT use a hierarchical query system to diagnose eight different major cancers from a single CT scan, outperforming specialized single-organ models by sharing feature representations across related pathologies.

Incorporating Domain Knowledge

Fine-grained classification is difficult because the visual differences between subtypes (e.g., two types of macular degeneration) are often subtler than the differences between a healthy and diseased organ. To overcome this, researchers are injecting clinical "priors" into their models.

Haihong E introduced KFWC, a knowledge-driven model for Wet-AMD. Instead of training the model on raw labels, they pre-trained it to recognize specific clinical lesion signs first. This "bottom-up" approach ensures the model focuses on the same features a human ophthalmologist would use to differentiate sub-conditions. Similarly, Zongru Shao deconstructed depression detection into 11 clinical sub-tasks based on the PHQ-9 questionnaire, forcing Large Language Models to reason through specific symptoms rather than relying on keyword biases.

From Detection to Virtual Biopsy

The ultimate goal of increasing granularity is to move toward "virtual biopsies"—non-invasive methods of determining a disease's molecular or aggressive profile. Yu‐Ying He demonstrated this by using radiomics to distinguish between aggressive and indolent subtypes of pediatric neuroblastoma. By extracting high-dimensional texture features that are invisible to the naked eye, the model provides a level of risk stratification that previously required invasive surgery.

As Hongyan Xu emphasizes, the transition from binary screening to multiclass differential diagnosis is what will move AI from a laboratory curiosity to a practical tool for precision medicine. By capturing the full spectrum of pathological variations, these models provide the specific etiological diagnoses necessary to guide personalized treatment.

Papers in this version

Showing 60 of 162 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_fine_grained_medical_diagnosis_beyond_bi_9bc1a7f2,
  title = {Fine-grained Medical Diagnosis Beyond Binary Classification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fine-grained-medical-diagnosis-beyond-binary-classification/rdr_9bc1a7f248114a68be3214f468aa5a73},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_fine_grained_medical_diagnosis_beyond_bi_9bc1a7f2_v1,
  title = {Fine-grained Medical Diagnosis Beyond Binary Classification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fine-grained-medical-diagnosis-beyond-binary-classification/rdr_9bc1a7f248114a68be3214f468aa5a73/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}