Source-linked AI summary
Anatomical and Diagnostic Bayesian Segmentation in Prostate MRI $-$Should Different Clinical Objectives Mandate Different Loss Functions?
Anindo Saha, Joeran Bosma, Jasper Linmans, Matin Hosseinzadeh, Henkjan Huisman
TL;DR
Anatomical and diagnostic prostate MRI segmentation pursue different clinical objectives, yet are often trained with similar losses. This study compares distribution-, region- and boundary-based losses in a probabilistic model across 200 ProstateX exams, finding that focal loss is especially suitable for diagnostic or panoptic segmentation, whereas most losses perform similarly for anatomical segmentation.
Problem
The paper asks whether anatomical organ quantification and diagnostic lesion detection require different loss functions despite being commonly treated as similar segmentation tasks.
Source: §1 · p. 1Method
The study compares CE, BCE, focal and Dice-plus-boundary losses in probabilistic anatomical and diagnostic prostate MRI segmentation across 200 ProstateX exams.
Source: §2 · p. 2Results
Distribution-based losses, especially focal loss, yield better calibration and substantially better lesion detection than Dice-plus-boundary loss, with a 15.7% increase in maximum detection sensitivity and 0.25 higher partial FROC area.
Source: §3 · p. 3Takeaways & Limitations
Distribution-based losses, particularly focal loss, are recommended for diagnostic or panoptic segmentation, while most compared losses perform similarly for anatomical segmentation.
Source: §3 · p. 4
Abstract
from arXiv · showhide
We hypothesize that probabilistic voxel-level classification of anatomy and malignancy in prostate MRI, although typically posed as near-identical segmentation tasks via U-Nets, require different loss functions for optimal performance due to inherent differences in their clinical objectives. We investigate distribution, region and boundary-based loss functions for both tasks across 200 patient exams from the publicly-available ProstateX dataset. For evaluation, we conduct a thorough comparative analysis of model predictions and calibration, measured with respect to multi-class volume segmentation of the prostate anatomy (whole-gland, transitional zone, peripheral zone), as well as, patient-level diagnosis and lesion-level detection of clinically significant prostate cancer. Notably, we find that distribution-based loss functions (in particular, focal loss) are well-suited for diagnostic or panoptic segmentation tasks such as lesion detection, primarily due to their implicit property of inducing better calibration. Meanwhile, (with the exception of focal loss) both distribution and region/boundary-based loss functions perform equally well for anatomical or semantic segmentation tasks, such as quantification of organ shape, size and boundaries.
1 Introduction
Prostate anatomy and clinically significant cancer have different clinical objectives, but are often trained and evaluated as near-identical segmentation tasks. The paper therefore motivates task-specific loss functions that reflect semantic organ quantification versus panoptic lesion detection.
- Anatomical segmentation supports prostate volume, boundary and geometry estimation for PSA density calculation and treatment planning, whereas cancer segmentation detects and characterizes malignant lesions for diagnosis and targeted biopsy.Source: §1 · p. 1
- Anatomical segmentation is a semantic single-object task, while clinically significant prostate cancer segmentation is a panoptic multi-object detection task.Source: §1 · p. 1
- Despite these differences, medical image studies commonly train both objectives with U-Nets and evaluate them similarly using Dice Similarity Coefficient.Source: §1 · p. 1
- Dice is appropriate for single-object quantification but cannot measure multi-object detection effectively, motivating task-specific losses during training.Source: §1 · p. 1
Segmentation Loss Functions
- Distribution-based losses minimize prediction–target distribution mismatch, whereas region or boundary-based losses optimize segmentation regions or contours; Dice-plus-boundary loss is designed to resist class imbalance.Source: §1 · p. 2
2 Experiments and Analysis
The study evaluates four segmentation losses in a probabilistic prostate MRI model using 200 ProstateX exams and matched experimental protocols. The Bayesian model uses uncertainty-aware inference, standardized preprocessing, and repeated cross-validation across anatomical and diagnostic tasks.
Materials
- The dataset comprised 200 prostate bpMRI exams from ProstateX with voxel-level delineations of WG, TZ, PZ and csPCa.Source: §2 · p. 2
Bayesian Segmentation Model
- A probabilistic deep attentive 3D U-Net used Monte-Carlo dropout to capture epistemic and aleatoric uncertainty during inference.Source: §2 · p. 2
Experiments
- Four losses—CE, BCE, focal loss and Dice-plus-boundary loss—were compared for anatomical segmentation of WG, TZ and PZ and diagnostic segmentation of csPCa.Class weights were task-specific, focal-loss γ was 2.00, and identical data splits were used across tasks.Source: §2 · p. 2
- Evaluation used three independent runs, five-fold cross-validation and the mean of 100 probabilistic executions per image, with metrics computed in 3D.Source: §2 · p. 2
Bayesian Segmentation Model
- At training time, the model combines KL divergence between prior and posterior distributions with segmentation loss, sampling z from Q to predict a mask.Source: Figure 1 · p. 2
3 Results and Discussion
Loss choice affects diagnostic and anatomical objectives differently: distribution-based losses, especially focal loss, improve calibration and lesion detection, while most losses achieve comparable anatomical segmentation quality. The results are assessed through calibration, DSC, patient-level AUROC and lesion-level FROC.
Calibration
- Distribution-based losses produce better calibration than Dice-plus-boundary loss, with focal loss showing the strongest calibration.The evaluation uses negative log-likelihood to assess calibration for anatomical and diagnostic segmentation.Source: §3 · p. 3, Figure 2 · p. 3
Anatomical Segmentation
- For anatomical segmentation, CE, BCE and Dice-plus-boundary loss achieve similar soft DSC, while focal loss performs worse with mean DSC of WG 70, TZ 64 and PZ 50.CE, BCE and DB achieve WG 81–89, TZ 82–85 and PZ 71–74 mean soft DSC.Source: §3 · p. 3
- With binarized argmax predictions, focal loss is only marginally below the other losses, whose anatomical mean DSC ranges are WG 88–90, TZ 84–86 and PZ 72–76.Source: §3 · p. 3
- The authors attribute focal loss’s anatomical decline to high predictive uncertainty near class boundaries, which limits precise contour definition.Source: §3 · p. 3
Diagnostic Segmentation
- Focal loss and other distribution-based losses substantially outperform Dice-plus-boundary loss for lesion detection, increasing maximum detection sensitivity by 15.7% and partial FROC area by 0.25.Patient-level diagnosis remained similar across losses at 76–78 mean AUROC.Source: §3 · p. 3, Figure 2 · p. 3
- Dice-plus-boundary loss can favor confident overlap of clear lesions, producing polarized predictions and a flatter FROC curve that misses difficult, low-confidence or small lesions.Source: §3 · p. 4
Broader Impact
The paper situates automated prostate cancer detection as a potential aid for radiologists facing difficult and variable MRI interpretation. It reports no foreseeable negative societal effects.
- Prostate cancer MRI interpretation is challenged by multifocality, morphological heterogeneity and resemblance to non-malignant conditions, contributing to inter-reader agreement below 50%.Source: Broader Impact · p. 4
- Automated detection algorithms could support consistent quantitative analysis, improve diagnostic accuracy and reduce unnecessary biopsies.Source: Broader Impact · p. 4
- The study reports no foreseeable negative societal effects.Source: Broader Impact · p. 4
A Appendix: Model Predictions
Figure 3 compares anatomical and diagnostic segmentations from four loss functions on one validation scan, pairing visual predictions with class-probability histograms.
- The figure compares WG, TZ, PZ, and csPCa segmentations produced with CE, BCE, FL, and DB.Source: Figure 3 · p. 6
- Predictions are overlaid on the T2W scan, with white lines marking the corresponding ground-truth annotations.Source: Figure 3 · p. 6
- Histograms below each case show class probabilities over predicted segments, indicating the loss functions’ implicit calibration.Source: Figure 3 · p. 6