Background Small and low-contrast liver tumors remain challenging targets for contrast-enhanced CT segmentation because of severe class imbalance and limited long-range contextual modeling in conventional CNN encoders. Methods We developed OF-TransUNet, a minimalist hybrid that differs from parameter-heavy TransUNet-style variants by inserting a single lightweight mid-level Conv-Transformer block at encoder stage 3 (×8 downsampling) within an otherwise unchanged 2D nnU-Net, together with an output-focused progressive unfreezing schedule intended to improve adaptation stability. This design was motivated by unstable and poorly reproducible optimization observed under immediate full fine-tuning in internal ablation experiments. A public dataset (n = 104) was used descriptively for architecture and schedule selection. The pre-specified primary endpoint was the external per-patient tumor Dice difference versus a standardized 2D nnU-Net baseline on an independent cohort (n = 42). A pre-defined secondary lesion-level analysis focused on medium-sized tumors (10–50 mm). Results On the external cohort, OF-TransUNet showed a numerically higher per-patient tumor Dice than nnU-Net (0.2788 ± 0.2575 vs 0.2400 ± 0.2426), with a mean paired difference of +0.0388 (95% CI 0.0029 to 0.0748). Because paired differences were non-normal, the pre-specified Wilcoxon signed-rank test was retained as the primary inferential analysis and was borderline (p = 0.0553); a supplementary paired t-test yielded nominal significance (p = 0.0347). In the pre-defined medium-lesion analysis (10–50 mm; n = 84), detection increased from 0.190 to 0.286 at the pre-specified 10% overlap threshold (McNemar exact p = 0.021; GEE OR 1.97, 95% CI 1.15–3.35). Post hoc sensitivity analyses using 5% and 15% overlap thresholds preserved the directional advantage. Relative to baseline, OF-TransUNet increased parameters by 8.4% and FLOPs by 18.3%, with no measurable latency penalty and only minimal memory increase. Conclusions In this single-center external validation cohort, a single mid-level Conv-Transformer insertion plus output-focused progressive unfreezing was associated with a numerically higher per-patient tumor Dice and a statistically supported improvement in the pre-defined medium-lesion detection analysis, while preserving a lightweight computational profile. Because the pre-specified non-parametric primary patient-level analysis was borderline and did not reach conventional significance, the Dice finding should be interpreted cautiously. Tumor HD95 was numerically higher in OF-TransUNet, indicating a possible recall-boundary trade-off that requires further boundary-focused evaluation. Overall, this minimally invasive modification supports further multi-center validation rather than definitive claims of superiority. A supplementary controlled benchmark on the public LiTS dataset, limited to a single pre-fixed validation fold because only 131 public cases have released annotations for local evaluation, provided directionally consistent contextual evidence under an identical pipeline.

Efficient transformer integration in nnU-Net for liver tumor segmentation: an external validation study

期刊论文
出版模式：

开放获取
发表日期：

2026年5月22日

查看全文

He Cao,
Li Tao,
Fan Li,
Xiaoping Pan

摘要

Background

Small and low-contrast liver tumors remain challenging targets for contrast-enhanced CT segmentation because of severe class imbalance and limited long-range contextual modeling in conventional CNN encoders.

Methods

We developed OF-TransUNet, a minimalist hybrid that differs from parameter-heavy TransUNet-style variants by inserting a single lightweight mid-level Conv-Transformer block at encoder stage 3 (×8 downsampling) within an otherwise unchanged 2D nnU-Net, together with an output-focused progressive unfreezing schedule intended to improve adaptation stability. This design was motivated by unstable and poorly reproducible optimization observed under immediate full fine-tuning in internal ablation experiments. A public dataset (n = 104) was used descriptively for architecture and schedule selection. The pre-specified primary endpoint was the external per-patient tumor Dice difference versus a standardized 2D nnU-Net baseline on an independent cohort (n = 42). A pre-defined secondary lesion-level analysis focused on medium-sized tumors (10–50 mm).

Results

On the external cohort, OF-TransUNet showed a numerically higher per-patient tumor Dice than nnU-Net (0.2788 ± 0.2575 vs 0.2400 ± 0.2426), with a mean paired difference of +0.0388 (95% CI 0.0029 to 0.0748). Because paired differences were non-normal, the pre-specified Wilcoxon signed-rank test was retained as the primary inferential analysis and was borderline (p = 0.0553); a supplementary paired t-test yielded nominal significance (p = 0.0347). In the pre-defined medium-lesion analysis (10–50 mm; n = 84), detection increased from 0.190 to 0.286 at the pre-specified 10% overlap threshold (McNemar exact p = 0.021; GEE OR 1.97, 95% CI 1.15–3.35). Post hoc sensitivity analyses using 5% and 15% overlap thresholds preserved the directional advantage. Relative to baseline, OF-TransUNet increased parameters by 8.4% and FLOPs by 18.3%, with no measurable latency penalty and only minimal memory increase.

Conclusions

In this single-center external validation cohort, a single mid-level Conv-Transformer insertion plus output-focused progressive unfreezing was associated with a numerically higher per-patient tumor Dice and a statistically supported improvement in the pre-defined medium-lesion detection analysis, while preserving a lightweight computational profile. Because the pre-specified non-parametric primary patient-level analysis was borderline and did not reach conventional significance, the Dice finding should be interpreted cautiously. Tumor HD95 was numerically higher in OF-TransUNet, indicating a possible recall-boundary trade-off that requires further boundary-focused evaluation. Overall, this minimally invasive modification supports further multi-center validation rather than definitive claims of superiority. A supplementary controlled benchmark on the public LiTS dataset, limited to a single pre-fixed validation fold because only 131 public cases have released annotations for local evaluation, provided directionally consistent contextual evidence under an identical pipeline.