Bidirectional Memory Improves Thoracic Contouring and Clinical Workflow

Thoracic radiotherapy contouring is a volumetric task performed one axial slice at a time, yet many segmentation systems still treat slices independently or carry context without accounting for whether anatomy has meaningfully changed. That mismatch is most visible at organ boundaries, in small low-contrast targets, and when a prediction must be trusted enough to enter a treatment-planning workflow. The authors introduce DAMM-Net++, a 2.5D auto-contouring system intended to make through-plane context, boundary precision, and uncertainty estimation part of the same clinical pipeline.

Core Contribution

DAMM-Net++ centers on anatomy-change-aware bidirectional selective state-space memory. Rather than propagating every slice feature equally through a CT volume, the model computes inter-slice changes and uses them to control what context should be accumulated in each through-plane direction. The premise is concrete: neighboring slices with stable anatomy should reinforce each other, while abrupt transitions should not be blurred into a single memory state.

The paper combines that memory with a boundary-aware decoder, multi-task outputs for segmentation, boundary and signed-distance prediction, a presence signal, and per-voxel uncertainty. This is more than an architectural accuracy claim. The intended use is triage: predictions with higher uncertainty can receive closer clinical review instead of being treated as equally reliable.

Technical Approach

Each of $T$ consecutive axial CT slices is encoded by a weight-shared ConvNeXt backbone. The Multi-Scale Inter-Slice Transition branch processes differences between slice features at several scales, with channel and spatial attention, to represent anatomy transitions. Those signals inform forward and backward selective state-space sweeps through the sequence. Their gated fusion gives the current slice access to contextual memory from both directions.

Figure 1 lays out this end-to-end design: slice encoding, transition-aware memory, and a memory-guided boundary-aware decoder that returns several prediction heads at full resolution. The architecture therefore remains 2.5D rather than using a fully volumetric encoder, while explicitly targeting the source of through-plane inconsistency that a slice-wise network misses.

The decoder uses memory-guided cross-attention to query the current representation against the accumulated context. The auxiliary boundary and signed-distance outputs make the loss sensitive to near-surface errors, where a small displacement can matter clinically even when overlap metrics remain high. The uncertainty output is also evaluated for calibration and is presented as a case-review aid, not as a substitute for clinician judgment.

Results and Analysis

The reported evaluation is substantial in scale: 2,146 patients from four centers, an independent external cohort of 112 patients, and a reader study with 17 radiation oncologists over 150 cases. Across the segmentation evaluation, DAMM-Net++ reports mean Dice of 0.955 and HD95 of 3.78 mm. The paper attributes its largest gains to low-contrast organs at risk and target volumes, precisely the structures for which continuity across slices should matter most.

The visual comparisons reinforce a narrower claim than headline averages alone. Figure 3 reports paired per-patient Dice distributions across nine thoracic structures and says DAMM-Net++ exceeds four baselines significantly on eight; Heart is the exception. Figure 4 shows axial, coronal, and sagittal reformats of one prediction, addressing whether apparently good axial masks remain coherent in the volume. These figures support the argument that the memory mechanism improves consistency, although they do not isolate which auxiliary head contributes most to each gain.

The reader study supplies the more consequential evidence. With AI assistance, contouring time fell by 75–80% across reader experience levels. Junior-reader mean IoU rose from 0.861 to 0.925, matching the unedited model according to the paper, while Figure 8 reports that all 17 readers improved in both mean IoU and contouring time. That is a practical result: the system appears to reduce editing work rather than merely produce a favorable offline Dice score.

External performance declined by less than 5% relative to internal evaluation, and the uncertainty calibration reportedly transferred without recalibration. The evidence supports use as an assistive contouring tool, particularly where experienced review remains available. It does not establish that faster contouring improves downstream treatment quality or patient outcomes, and the reported validation remains confined to thoracic radiotherapy structures.