Interpretable predictions from whole-body FDG-PET/CT using parameters associated with clinical outcome

Sambit Tarai; Elin Lundström; Nouman Ahmad; Robin Strand; Håkan Ahlström; Joel Kullberg

doi:10.1038/s43856-026-01567-w

Interpretable predictions from whole-body FDG-PET/CT using parameters associated with clinical outcome

Commun Med (Lond). 2026 Apr 20;6(1):232. doi: 10.1038/s43856-026-01567-w.

Authors

Sambit Tarai¹, Elin Lundström², Nouman Ahmad², Robin Strand³, Håkan Ahlström^{2

4}, Joel Kullberg^{2

4

5}

Affiliations

¹ Radiology, Department of Surgical Sciences, Uppsala University, Uppsala, Sweden. sambit.tarai@uu.se.
² Radiology, Department of Surgical Sciences, Uppsala University, Uppsala, Sweden.
³ Department of Information Technology, Uppsala University, Uppsala, Sweden.
⁴ Antaros Medical AB, Mölndal, Sweden.
⁵ Department of Surgical Sciences, SciLifeLab, Uppsala University, Uppsala, Sweden.

Abstract

Background: Accurate prediction of clinical outcomes is challenging yet important for patient care. The aim of the study was to evaluate a deep learning-based methodology using tissue-wise information, as a proof of concept, for predicting parameters known to be associated with clinical outcomes.

Methods: We utilized the publicly available autoPET cohort, consisting of 1014 FDG-PET/CT examinations. Tissue-wise projections were extracted, representing specific tissues (bone, lean tissue, adipose tissue, and air) at different angles. A deep regression and classification framework was trained to predict total metabolic tumor volume (TMTV), lesion count, patient age, sex, and diagnosis status (cancer vs. no cancer). Saliency analysis was performed to identify image regions contributing most to each prediction.

Results: Here we show that the best model predicts TMTV (MAE = 77 ml; R² = 0.84; (p <0.05)) and lesion count (MAE = 5.18, R² = 0.90), when using all tissue-wise projections. Age prediction improves when multiple projection angles are included (MAE = 6.57 years; R² = 0.70 (p <0.05)). The model also predicts sex (AUC = 1.00 (p <0.05)) and diagnosis status with high accuracy (AUC = 0.95 (p <0.05)).

Conclusions: This proof-of-concept study demonstrates that tissue-wise projections can be used for efficient and automated prediction of parameters related to clinical outcomes, highlighting their potential for future prediction of clinical outcomes.

Plain language summary

Cancer is a leading global cause of death. Physicians use images obtained from medical scans to detect cancer, guide treatment, and monitor outcomes. Artificial intelligence (AI) can support this process, but its decision making is often difficult to interpret. We have developed an AI model that first converts 3D scans into 2D projections. The AI model then utilizes these projections to predict key health parameters related to outcome, such as tumor volume, age, sex and cancer status. Our results demonstrate that the AI is accurate while also providing the ability to visualize the specific image regions it identifies as important for each prediction task. This transparency enhances trust in AI and demonstrates its potential to support personalized cancer care.