Source-linked AI summary
FD-AA: A Lightweight Focal-Diffuse And Attenuation-Aware Head for Incidental Abdominal Abnormality Detection in Chest CT
Haoyan Ding, Kritika Iyer, Halid Yerebakan, Zhenyu Bu, Chushu Shen, Peiyu Duan, Xinyuan Zheng, Sepehr Farhand, Xueqi Guo, Chaowei Wu, Yoshihisa Shinagawa, Gerardo Hermosillo Valadez
TL;DR
Routine chest CT captures upper-abdominal organs where incidental abnormalities may offer added clinical value, but global representations can dilute localized findings and visual features omit HU intensities. FD-AA combines focal and diffuse organ evidence with attenuation cues in a lightweight head, achieving the highest macro AUC, AP, and F1 on internal and external cohorts with frozen Pillar-0.
Problem
Chest CT captures abdominal organs with incidental abnormalities, but scan-level global representations can dilute localized findings and learned visual features do not preserve quantitative HU intensities.
Source: §1 · p. 1Method
FD-AA combines sparse focal and diffuse organ evidence with high- and low-attenuation cues in a lightweight head for frozen 3-D encoders.
Source: §1 · p. 1Results
FD-AA with frozen Pillar-0 achieved the highest macro AUC, AP, and F1 on both the CT-RATE test set and external RAD-ChestCT cohort, and improved AUC and AP across four evaluated encoders.
Source: §3.4 · p. 4, §3.3 · p. 4Takeaways & Limitations
The results support combining focal-diffuse aggregation with explicit attenuation cues while preserving spatial and quantitative CT evidence without training a separate encoder.
Source: §4 · p. 4FD-AA relies on organ masks at inference, and abdominal organs may be only partially captured in chest CT.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Routine chest CT captures upper-abdominal structures that may contain clinically relevant incidental abnormalities. Detecting these findings requires feature extraction from organs with different spatial extents and attenuation patterns. We propose FD-AA, a lightweight organ-aware classification head adaptable for frozen 3-D CT encoders. Within each organ, an attenuation-aware module preserves sparse focal evidence, while masked generalized-mean pooling captures diffuse anomaly patterns. In seven abdominal organs, FD-AA with Pillar-0 achieved state-of-the-art (SOTA) performance in both the CT-RATE test set (AUC = 0.798) and the external RAD-ChestCT dataset (AUC = 0.713). More specifically, FD-AA improved macro AUC/AP from 0.763/0.346 to 0.798/0.405 over direct classification using frozen Pillar-0 only (p = 0.034/0.016). Such performance gain generalizes across multiple frozen encoders (AUC improvement on MedicalNet +9.8%, CT-CLIP +14.7%, ResNet +3.7%), demonstrating the effectiveness of FD-AA across different feature representations. These results support the effectiveness of integrating focal-diffuse aggregation with explicit HU evidence for incidental abdominal abnormality detection.
1. INTRODUCTION
Routine chest CT captures upper-abdominal organs where incidental abnormalities may offer added clinical value. FD-AA addresses limitations of global aggregation and visual-only features with focal-diffuse organ evidence and HU-derived cues, and is reported to achieve state-of-the-art results on two datasets.
- Routine chest CT often captures upper-abdominal structures where incidental abnormalities may be detected, offering additional clinical value without extra scanning or radiation exposure.Source: §1 · p. 1
- Scan-level global representations can dilute localized organ abnormalities, while learned visual features alone do not preserve quantitative HU intensities.Abdominal abnormalities vary in spatial extent and appearance, and organs may be only partially captured in chest CT.Source: §1 · p. 1
- FD-AA with frozen Pillar-0 established state-of-the-art performance on CT-RATE and RAD-ChestCT, outperforming prior supervised baselines and competitive CT-VLMs.Source: §1 · p. 1
- FD-AA combines sparse focal and diffuse organ evidence with an attenuation-aware module that represents high- and low-attenuation changes while discounting extra-organ confounders.Source: §1 · p. 1
2. METHOD
FD-AA uses frozen-encoder features and organ masks to model focal and diffuse evidence across seven abdominal targets. Its attenuation-aware module adds HU-derived evidence, and staged training combines the resulting features while keeping the encoder frozen.
2.1. Problem Definition and Preliminaries
- A frozen 3-D encoder produces spatial tokens and a global embedding for seven abdominal targets, with TotalSegmentator masks aligned to the token grid to define organ support.The targets are abdominal tissue, adrenal gland, aorta, gallbladder, kidney, liver, and spleen.Source: §2.1 · pp. 1–2, Figure 1 · p. 2
2.2. Focal-Diffuse Organ-Aware Head
- The organ-aware head combines top-5% focal token pooling and masked GeM for diffuse evidence, then uses global context and organ-specific classifiers and mixing coefficients.The GeM exponent is learned, bounded between 1 and 6, and initialized at 3.Source: §2.2 · p. 2
2.3. Attenuation-Aware Module
- Token-aligned HU descriptors capture high- and low-attenuation evidence, contrast within-organ high HU against an extra-organ ring, and guide visual-token attention through a residual correction to focal features.High-density thresholds are 130, 200, and 400 HU; low-density descriptors use 20 and 0 HU thresholds plus offsets below the within-organ median.Source: §2.3 · pp. 2–3
2.4. Training Objectives
- Training keeps the encoder frozen, uses organ-weighted abnormality and report-derived density-finding supervision, and then optimizes residual fusion with fused-logit loss and an L2 correction penalty.Attention losses encourage within-organ focus and align attention with label-selected evidence; these labels are not required at inference.Source: §2.4 · p. 3
3. EXPERIMENTS AND RESULTS
FD-AA with a frozen Pillar-0 encoder achieves the strongest reported results across the internal CT-RATE test set and external RAD-ChestCT cohort. Across four frozen encoders, the head consistently improves AUC and AP; the evaluation uses seven abdominal targets and specified training settings.
3.1. Datasets and Evaluation
- CT-RATE contributes 50,178 scans split into training, validation, and test sets, while external validation uses all 3,630 RAD-ChestCT scans mapped to the same seven targets.The CT-RATE splits have no patient overlap; macro AUC and AP are primary metrics due to class imbalance.Source: §3.1 · p. 3
3.2. Implementation Details
- Experiments reuse precomputed tokens from a frozen Pillar-0 encoder; the focal pathway retains the top 5% of organ tokens and GeM pooling starts at p = 3.The focal-diffuse head and attenuation-aware module use AdamW with a learning rate of 3 × 10^-4; the residual-fusion adapter uses 5 × 10^-4.Source: §3.2 · p. 3
3.4. Comparison with State-of-the-Art Methods
- FD-AA achieves the highest macro AUC, AP, and F1 on both CT-RATE and RAD-ChestCT, and the highest label-wise accuracy on CT-RATE.All results use the FD-AA head with a frozen Pillar-0 encoder.Source: Table 1 · p. 4, §3.4 · p. 4
3.3. Evaluation Across Image Encoders
- Across four frozen encoders, FD-AA consistently improves AUC and AP, with Pillar-0 achieving the strongest overall performance.The compared configurations test global-aware, organ-aware, focal-diffuse, and attenuation-aware heads.Source: Table 2 · p. 4, §3.3 · p. 4
4. DISCUSSION AND CONCLUSION
The study presents a lightweight head that combines focal and diffuse organ evidence with attenuation-aware modeling for incidental abdominal abnormality detection. It reports improved internal and external ranking metrics with a frozen 3-D foundation encoder, while noting data, localization-validation, and inference-mask constraints.
4. DISCUSSION AND CONCLUSION
- Report-derived labels may be noisy or omit subtle findings, lesion-level annotations are absent, and inference relies on organ masks despite partial organ coverage in chest CT.Future work will explore mask-free inference and partial organ coverage.Source: §4 · p. 4
- The lightweight head combines sparse focal and diffuse organ evidence with attenuation-aware modeling to improve internal and external ranking metrics using a frozen 3-D foundation encoder.The approach preserves spatial and quantitative CT evidence without training a separate encoder.Source: §4 · p. 4