Dual-Phase Anatomical Prior Learning for Liver Lesion Segmentation on Non-Contrast CT

Automated liver lesion segmentation on computed tomography (CT) is important for quantitative assessment, treatment planning, and longitudinal evaluation [6]. Most existing liver lesion segmentation benchmarks and methods focus on contrast-enhanced CT [2], where contrast administration improves lesion conspicuity and facilitates differentiation from the surrounding liver parenchyma. However, contrast-enhanced imaging is not always feasible, for example when contrast administration is contraindicated [7] or contrast-enhanced acquisitions are unavailable [5]. In such settings, non-contrast CT (NCCT) becomes an important alternative, but liver lesion segmentation is substantially more challenging because lesions often exhibit subtle attenuation differences and poorly defined boundaries [4]. This challenge motivates the use of complementary information from contrastenhanced phases during model training. Existing multi-phase CT studies have exploited such information through multi-channel inputs, phase-specific feature fusion, attention mechanisms, and spatial cross-phase aggregation [10, 11, 1]. However, these approaches commonly assume that multiple phases are also available during inference. The recently introduced TriALS challenge [4] considers a different setting by providing paired non-contrast, arterial, portal venous, and delayed CT images during training while restricting inference to NCCT alone. Anatomical prior learning provides another way to complement weak local appearance in 2D segmentation. The Anatomical Codebook framework [8] learns position-aware anatomical prototypes from auxiliary CT slices through a shared learnable codebook and uses the resulting prototype-weighted representations to enrich 2D segmentation features. This formulation enables anatomical context learned during training to be accessed from a single image at inference. Building on this codebook-based anatomical prior formulation, we investigate its application to NCCT liver lesion segmentation under the TriALS setting, with particular focus on whether anatomical priors learned with registered portal venous CT provide complementary context for NCCT liver lesion segmentation. Specifically, we extend the auxiliary prior-learning setting to jointly use NCCT and portal venous CT registered to the NCCT space. The two phases contribute to a shared anatomical prior space, while segmentation supervision remains entirely NCCT-based. This dual-phase setting combines target-domain NCCT information with anatomical cues learned from the more conspicuous portal venous phase without changing the deployment requirement. Portal venous CT is therefore used only during training, while inference requires NCCT alone. The main contributions of this work are: – We develop a dual-phase prior-learning framework in which NCCT and registered portal venous CT jointly contribute to a shared anatomical codebook during training, while inference remains NCCT-only. – Experiments on the TriALS dataset show that the proposed framework improves NCCT liver lesion segmentation over the U-Net baseline, achieving gains of 7.23%p in Dice and 6.26%p in IoU.

The description of the methodology is explained as the following subsections.

We build on the codebook-based anatomical prior formulation of Anatomical Codebook [8], which learns position-aware anatomical prototypes and uses prototype-weighted representations to enrich 2D segmentation features. In our framework, this formulation is implemented at the local patch level and integrated with a 2D U-Net for NCCT liver lesion segmentation. As illustrated in Fig. 1, training contains two complementary components: an NCCT segmentation stream and a dual-phase auxiliary prior-learning stream. The segmentation stream is supervised only by NCCT images and their lesion annotations, whereas auxiliary NCCT and registered portal venous slices jointly contribute to the same anatomical codebook. The learned codebook prior is subsequently fused with the U-Net features for segmentation. Portal venous CT is required only during training; inference uses NCCT alone.

Let x^N and x^V denote NCCT and portal venous slices, respectively, where the portal venous volume is registered to the NCCT coordinate space. For anatomical prior learning, auxiliary slices from each phase are partitioned into a P × P grid of local patches. A shared patch encoder E_p(·) maps patch t from phase m ∈ {N, V } to a feature representation (Equation 1), where (Notation 1) denotes the t-th patch from phase m, with N and V representing NCCT and portal venous CT, respectively, and (Notation 2) is the corresponding patch embedding. Following the codebook-based representation principle in [8], each patch embedding is compared with a shared learnable codebook (Equation 2), where K is the number of codewords and e_k denotes the k-th codeword. For patch t from phase m ∈ {N, V }, the soft assignment probability to codeword k is computed as (Equation 3), where (Notation 3) denotes the assignment probability of patch t to codeword k, s(·, ·) is the feature-similarity function, τ is the temperature parameter, and j indexes all codewords in the softmax normalization. The anatomical prior representation for patch t is then obtained as the weighted combination of all codewords: (Equation 4), where (Notation 4) denotes the codebook-derived anatomical prior representation of patch t from phase m. Both NCCT and venous auxiliary patches use the same patch encoder and anatomical codebook. Therefore, rather than learning two phase-specific prior spaces, samples from both phases jointly shape a common set of anatomical prototypes. Because both phases share the same codebook and position-regression objective, NCCT and registered portal venous patches at corresponding anatomical locations are trained to predict the same spatial coordinates. Through this supervision, the shared codebook is encouraged to form prototypes that capture features informative for distinguishing anatomical positions, thereby establishing a position-aware anatomical prior space jointly shaped by both phases. Entropy regularization further encourages balanced codebook utilization. For position supervision, the codebook-derived representation (Notation 4) is mapped by a shared position regressor R(·) to a three-dimensional position representation: (Equation 5). Rather than directly regressing absolute coordinates, we supervise the relative physical offset between foreground patch pairs. For a pair (u, v), let (Equation 6) where (Notation 5) denotes the physical 3D location of patch t. The position loss for phase m is then (Equation 7), where (Notation 6) denotes the set of foreground patch pairs, σ(·) is the sigmoid function, and α and β are scaling coefficients. The key adaptation in this work lies in the source of auxiliary prior information. NCCT provides target-domain appearance, whereas the registered portal venous phase provides an additional contrast-enhanced observation of the same anatomy. By jointly using both phases during codebook learning, the framework can incorporate complementary phase information without requiring portal venous CT as an input to the segmentation model at deployment.

The segmentation branch remains entirely NCCT-based. Given a labeled NCCT slice x^N with combined mask y^N , a 2D U-Net extracts multi-scale image features. The same NCCT slice also queries the learned anatomical codebook to obtain codebook-derived prior representations. The patch-level priors are spatially rearranged and mapped to the corresponding U-Net feature resolutions, and are fused with the segmentation features at multiple scales before decoding. This allows the segmentation network to combine local NCCT appearance with the anatomical context encoded by the shared codebook. The final prediction is (Equation 8), where F_seg denotes the U-Net-based segmentation network equipped with multi-scale anatomical prior fusion using the shared codebook C. The segmentation loss L_seg is computed only against the combined maske for NCCT. Neither portal venous images nor venous lesion annotations are used for segmentation supervision.

For each auxiliary phase m ∈ {N, V }, the anatomical prior objective combines the position supervision defined in Sec. 2.2 with entropy regularization following: (Equation 9), where λ_pos = 1.0 and λ_ent = 0.02 in all experiments. The smaller entropy weight regularizes codebook utilization without overwhelming the position-supervision objective. To maintain a comparable auxiliary-loss magnitude when introducing the second phase, the NCCT and venous prior objectives are averaged: (Equation 10). The complete training objective is: (Equation 11). At inference, the NCCT slice accesses position-aware anatomical priors jointly learned from NCCT and portal venous CT through the shared codebook, thereby improving NCCT-only lesion segmentation without requiring portal venous input.

The description of the experiments is explained as the following subsections.

Experiments were conducted on the TriALS dataset [4], which provides paired non-contrast, arterial, portal venous, and delayed-phase CT scans with lesion annotations. We used the 80 publicly available cases and retained only the NCCT and portal venous phases. The data were split at the patient level into 56 training, 12 validation, and 12 test cases to prevent slice-level leakage. Portal venous volumes were spatially registered to the corresponding NCCT volumes before constructing paired axial slices. CT intensities were clipped to [−200, 300] HU and normalized to [0, 1]. All images were resized to 512×512. All axial slices, including background-only slices, were retained, resulting in 28,730 training, 5,671 validation, and 5,718 test slices. A 2D U-Net [9] was used as the segmentation backbone for all experiments. For the anatomical codebook models, we used a 4 × 4 patch grid and a shared codebook with K = 1024 prototypes. The codebook formulation follows the anatomical prototype-learning principle of [8], while our implementation operates on local patches and uses NCCT and registered portal venous CT as auxiliary training inputs. All models were trained from scratch using the same data split, preprocessing, and optimization settings to ensure a controlled comparison. The best checkpoint for each experiment was selected according to the validation Dice Similarity Coefficient (DSC). Performance was evaluated on the held-out NCCT test set using DSC, Intersection over Union (IoU), precision, and recall.

As shown in Table 1, the ablation results reveal a complementary effect between dual-phase learning and position supervision. The dual-phase codebook without position supervision achieved high recall (0.6324) but low precision (0.3005), indicating substantial over-segmentation. Conversely, the NCCT-only position prior produced the highest precision (0.5618) but substantially reduced recall (0.2929), suggesting overly conservative predictions. Combining both components alleviated these complementary failure modes, yielding a more balanced precision–recall trade-off and the best overall DSC (0.5160) and IoU (0.3477). Adding position supervision to the dual-phase codebook improved DSC by 10.86%p, while adding portal venous information to the NCCT position prior improved DSC by 13.18%p. Overall, these results suggest that the benefit arises from jointly exploiting phase-complementary appearance information and position-guided anatomical priors rather than from either component alone. Recent results from the official TriALS 2025 challenge are summarized in Table 2. The challenge evaluated six participating methods on three hidden test cohorts under the combined-annotation paradigm, with the best per-case Dice ranging from 0.471 to 0.651 across centres [4]. These results provide useful context for the current state of NCCT liver lesion segmentation. However, direct numerical comparison with our results is not appropriate because the official challenge test data and annotations are not publicly available, whereas our experiments use a patient-level held-out subset of the 80 publicly released cases. Limitations and future directions. Despite the improvement over the U-Net baseline, the absolute segmentation performance remains moderate. One major challenge is the inherently low lesion conspicuity on NCCT, particularly for small or contrast-dependent lesions. The official TriALS benchmark similarly reports substantially lower performance on NCCT and severe difficulty for very small lesions [4]. Methodologically, our 2D formulation does not explicitly exploit through-plane context, while the fixed 4 × 4 patch grid may provide relatively coarse spatial priors for small or irregular lesions. In addition, the current codebook is primarily organized by anatomical position rather than lesion-specific appearance, which may limit its ability to distinguish subtle lesions from normal parenchyma. Residual registration differences between NCCT and portal venous CT may also introduce noise into dual-phase prior learning. Future work will investigate multi-scale or adaptive patch priors, 2.5D/3D contextual modeling, lesion-aware codebook objectives, and more robust cross-phase representation alignment while preserving NCCT-only inference.

We proposed a dual-phase position-guided anatomical prior learning framework for liver lesion segmentation on non-contrast CT (NCCT). The proposed method jointly learns a shared anatomical codebook from NCCT and spatially registered portal venous CT, while position regression encourages the codebook prototypes to capture features informative for anatomical location. Portal venous CT is used only as auxiliary information during training, whereas inference requires only a single NCCT image. On the TriALS dataset, the proposed method achieved a DSC of 0.5160 and an IoU of 0.3477, improving the U-Net baseline by 7.23%p and 6.26%p, respectively. The ablation results further showed that combining dual-phase learning with position supervision provided the most balanced precision–recall performance and the best overall segmentation accuracy. These results demonstrate that complementary information learned from contrast-enhanced portal venous CT can be effectively incorporated through position-guided anatomical priors to improve NCCT liver lesion segmentation without requiring additional contrast-enhanced input at inference.

This work was partly supported by the National Research Foundation of Korea(NRF) funded by the Korea government(MSIT)(RS-2024-00360226, 50%) and Institute of Information & Communications Technology Planning & Evaluation(IITP)-ITRC(Information Technology Research Center) grant funded by the Korea government(MSIT)(IITP-2026-RS-2024-00437718, 50%)