Source-linked AI summary
Parameter-Efficient pretrained-CT-to-MRI Transfer for Rectal Cancer Segmentation: Performance-Calibration Trade-offs
Aneesh Rangnekar, Jorge Tapias Gomez, Joseph O Deasy, Harini Veeraraghavan
TL;DR
Rectal-cancer MRI segmentation for adaptive radiotherapy needs accurate contours alongside efficient computation and trustworthy uncertainty. The paper benchmarks CT-pretrained Swin V2 transfer using decoder compression, LoRA, and a four-member ensemble, finding efficiency and calibration gains within a retrospective single-institution cohort.
Problem
Rectal MRI segmentation must provide geometric accuracy, practical computational cost, and calibrated uncertainty for adaptive radiotherapy, but their interplay had not been benchmarked holistically.
Method
A Swin V2 encoder pretrained on 10,444 public CT volumes was adapted to T2-weighted MRI through full fine-tuning, efficient decoding, LoRA, and a four-member LoRA-decoder ensemble.
Results
SWIFTe reduced total parameters by 70.1% versus SWIFT while increasing detection from 89.9% to 93.9%, and SWIFTe-LDE4 achieved ECE 0.217 and Brier score 0.222 after temperature scaling.
Takeaways & Limitations
Within the retrospective cohort, parameter-efficient adaptations preserved segmentation performance, while LDE4 improved calibration relative to the other evaluated configurations.
Takeaways & Limitations
The study used one tumor type from a single institution and GE scanners, while uncertainty signals were not quantitatively validated against manual correction effort or dosimetric impact.
Abstract
from arXiv · showhide
Accurate rectal cancer segmentation from magnetic resonance imaging (MRI) is essential for adaptive radiotherapy and tumor response assessment, but deployment also requires computational efficiency and informative, calibrated uncertainty estimates. We therefore introduce SWIFT, a SWin pretrained model wIth parameter-eFficient and Tumor-aware fine-tuning for rectal cancer segmentation. A Swin V2 encoder pretrained on 10,444 public 3D CT volumes using a DINOv2-style objective was adapted to T2-weighted MRI through four cumulative configurations: full fine-tuning (SWIFT), decoder compression (SWIFTe), low-rank adaptation (SWIFTe-LoRA), and a four-member LoRA-decoder ensemble (SWIFTe-LDE4). Geometric accuracy, tumor detection, radiomic agreement, and probability calibration were evaluated on a held-out 247-case test set from a single-institution cohort acquired using 1.5 or 3 Tesla GE scanners. Compared with SWIFT, SWIFTe reduced total parameters by 70.1% (from 72.8M to 21.8M) and increased tumor detection rate from 89.9% to 93.9%, while achieving a slightly lower median surface DSC (0.61 versus 0.62) and improved radiomic agreement. In a separate SWIFTe ablation, removing tumor-aware augmentation reduced detection from 93.9% to 89.9% but increased surface DSC from 0.61 to 0.64, demonstrating a detection-boundary-agreement trade-off. SWIFTe-LoRA used 14.6% of SWIFTe's trainable parameters while retaining similar segmentation performance. SWIFTe-LDE4 achieved the lowest calibration errors among the four configurations after temperature scaling (expected calibration error, 0.217; Brier score, 0.222), although the absolute expected calibration error indicates residual miscalibration. Similar efficiency-calibration patterns were observed using the public VoCo checkpoint, supporting robustness across pretrained initializations rather than external clinical generalizability.
1 Introduction
The paper frames rectal-tumor MRI segmentation for adaptive radiotherapy as requiring accuracy, computational efficiency, and calibrated uncertainty. It introduces a cumulative benchmark of decoder compression, LoRA adaptation, and ensemble calibration in a pretrained transfer pipeline.
- Daily rectal tumor contouring is a major bottleneck in MR-guided adaptive radiotherapy, which aims to adapt treatment while limiting exposure to nearby radiosensitive organs.
- DL models for time-sensitive adaptive radiotherapy must balance geometric accuracy, computational cost, and calibrated uncertainty for guiding user intervention.
- Prior work addressed accuracy, uncertainty reliability, and computational bottlenecks separately, leaving their interplay underexplored for reliable tumor segmentation.
- The benchmark applies four cumulative configurations: full fine-tuning with SWIFT, an efficient decoder in SWIFTe, LoRA adaptation in SWIFTe-LoRA, and ensemble branches in SWIFTe-LDE4.SWIFT uses a Swin transformer V2 encoder and retains the standard Swin UNETR V2 decoder for full-adaptation reference.
- The study evaluates whether reducing decoder capacity, freezing pretrained encoders with LoRA, and adding member-specific adapter-decoder pairs alter efficiency, segmentation, and uncertainty calibration.
2 Methodology
The methodology transfers a CT-pretrained Swin V2 encoder to T2-weighted MRI through progressively parameter-efficient decoder and adapter configurations. Evaluation combines segmentation, detection, radiomic agreement, uncertainty, and calibration analyses on patient-specific splits.
- Data and preprocessing: 416 T2-weighted MRIs were split into 136 training, 33 validation, and 247 testing cases after isotropic resampling, reorientation, and intensity clipping.The cohort used 1.5 or 3 Tesla GE scanners with phased-array coils.
- Transfer pipeline: SWIFT fine-tunes a 3D Swin UNETR V2 end-to-end after self-supervised DINOv2-style pretraining on 10,444 heterogeneous public CT volumes.The network uses a hierarchical Swin V2 encoder, multi-scale skip connections, feature size 48, and two-channel background/tumor output.
- Efficient decoder: SWIFTe isolates decoder compression by replacing the Swin UNETR decoder with EffiDec3D while retaining the CT-pretrained encoder and downstream protocol.EffiDec3D uses fixed 48-channel decoder stages and reconstructs reduced-resolution logits before trilinear upsampling.
- Parameter-efficient adaptation: SWIFTe-LoRA freezes the pretrained encoder and trains rank-4 adapters in attention projections jointly with the EffiDec3D decoder.SWIFTe-LDE4 adds four member-specific LoRA-adapter and decoder pairs sharing the frozen encoder; probabilities are averaged and disagreement measures uncertainty.
- Training and uncertainty: Tumor-aware augmentation locally transforms foreground tumor voxels during training, while case-level uncertainty uses predictive entropy and, for LDE4, ensemble variance and disagreement.The augmentation leaves surrounding anatomy unchanged.
- Evaluation: Segmentation evaluation reports detection rate and 2 mm-tolerance surface DSC across all 247 test cases, with zero sDSC assigned to model-specific misses.Calibration uses validation-set temperature scaling and reports reliability diagrams, ECE, and Brier score; radiomic agreement uses Lin’s concordance correlation coefficient.
3 Results
Across the SWIFT progression, efficiency and calibration improved through decoder compression, LoRA adaptation, and ensembling, while segmentation overlap and detection revealed explicit trade-offs. Radiomic agreement and efficiency–calibration trends were also assessed across configurations and pretrained initializations.
- Segmentation and efficiency: 70.1% fewer total parameters reduced SWIFT from 72.8M to 21.8M in SWIFTe, while detection increased to 93.9% and median sDSC was 0.61.SWIFT had 89.9% detection and median sDSC of 0.62.
- Calibration: 0.217 ECE and 0.222 Brier score were the lowest calibration errors for SWIFTe-LDE4, while median sDSC remained similar to SWIFTe-LoRA.SWIFTe-LDE4 reduced ECE from 0.245 to 0.217 and Brier score from 0.244 to 0.222; its median sDSC was 0.60 versus 0.59.
- Radiomic agreement: 0.69 median manual-versus-automated CCC was achieved by both SWIFTe-LoRA and SWIFTe-LDE4 across 288 radiomic features, compared with 0.62 for SWIFT.The fraction of features with CCC ≥0.75 was 39% for SWIFTe-LoRA and 38% for SWIFTe-LDE4, versus 30% for SWIFT.
- Parameter-efficient adaptation: 3.2M trainable parameters represented 14.6% of SWIFTe’s burden for SWIFTe-LoRA, which retained similar median sDSC but had lower detection.SWIFTe-LoRA median sDSC was 0.59 versus 0.61 for SWIFTe, while detection was 89.1% versus 93.9%.
- Initialization robustness: The VoCo progression reproduced the overall pattern: efficient decoding reduced model size, LoRA reduced trainable burden, and LDE4 improved calibration with broadly similar sDSC.The VoCo experiment tested trend consistency across pretrained initializations rather than directly comparing pretraining methods.
- Ablation: Removing tumor-aware augmentation lowered SWIFTe detection from 93.9% to 89.9% but increased median sDSC from 0.61 to 0.64.The ablation therefore exposed a detection–boundary-agreement trade-off.
4 Discussion and conclusions
Parameter-efficient CT-pretrained adaptations preserved segmentation performance, while SWIFTe-LDE4 improved relative calibration. Radiomic agreement and uncertainty analyses revealed behavior not captured by overlap alone, but clinical generalizability and prospective uncertainty-guided review remain unverified.
- 70.1% fewer total parameters maintained segmentation performance with the efficient decoder, while enhancing tumor detection.LoRA adapter configurations achieved detection performance comparable to full fine-tuning.
- SWIFTe-LDE4 improved calibration relative to the other evaluated SWIFT configurations without substantially degrading segmentation accuracy.
- Texture radiomic features were more reproducible for SWIFTe and LoRA-adapted configurations, whereas shape features remained less reproducible across configurations.This pattern was consistent with modest geometric agreement against manual contours.
- Higher entropy and, for SWIFTe-LDE4, higher ensemble disagreement generally identified lower-confidence cases for qualitative review prioritization rather than validated spatial error localization.
- Robustness across institutions, vendors, and imaging protocols remains unverified because evaluation used one tumor type and a single-institution cohort acquired exclusively on GE scanners.The uncertainty signals were also not quantitatively validated against manual correction effort or dosimetric impact.