The work targets a medically important but computationally difficult structure. Pulmonary fissures divide the lungs into distinct lobes, and their integrity can influence assessment of collateral ventilation before bronchoscopic lung-volume reduction procedures involving endobronchial valves. Incomplete fissures can permit gas and inflammation to move between adjacent lung regions, undermining the intended effect of valve-based treatment. Fissure detection also has roles in lesion localization and analysis of pulmonary diseases. Yet fissures often appear as faint, curved sheets with only one- or two-pixel-wide boundaries in CT imagery. In diseased lungs they may be deformed, interrupted or physically difficult to distinguish, making a tool that merely traces visible edges potentially unreliable.

Earlier computational approaches pursued this task with line-enhancement filters, ridge operators, derivative-based methods, anatomical atlases and watershed transformations. The study identifies two limitations associated with many deep-learning systems. First, deep networks commonly display a spectral bias, converging faster on low-frequency structures such as overall organ shapes than on high-frequency discontinuities such as fissure edges. The authors also cite evidence that multi-head self-attention can function as a spatial low-pass filter, attenuating fine signals and contributing to oversmoothing in deeper layers. Second, conventional Transformer attention has quadratic computational cost with respect to the number of tokens, creating a bottleneck for high-resolution three-dimensional volumes and encouraging compromises between spatial detail and long-range context.

CFT-Triplet addresses these constraints with a block that combines three kinds of information rather than simply stacking separate CNN and Transformer modules. Its Textural Branch uses cascaded residual convolutional units to retain local spatial patterns and edge information. Its Contextual Branch employs bidirectional Mamba2, a state-space model designed to model long-range dependencies with linear computational complexity. The Mamba2 configuration processes features in forward and reverse sequences and combines them through element-wise addition, allowing each direction to contribute to a global representation of anatomical continuity. Rather than operating as three independent branches, the Contextual and Frequency components are connected in series, with spectral refinement applied after contextual modeling.

The Frequency Branch is intended specifically to recover boundary information that may be weakened during global feature aggregation. Each directional feature stream passes through a learnable weighted fast Fourier transform block for spectral filtering and a multilayer perceptron for channel mixing. A learnable channel-wise parameter then balances the refined forward and backward features. The authors describe the effect as adaptive bandpass filtering: their feature-map analysis indicated that mid-to-high-frequency components were amplified without introducing extreme high-frequency noise. A separate Gate-Net fusion layer combines this sharpened global representation with the parallel Textural Branch. The gate, generated by two three-dimensional convolutional layers, a ReLU activation and a sigmoid function, spatially recalibrates the two streams so that the network can favor anatomical continuity in homogeneous regions and fine detail near boundaries.

The network was implemented as a hierarchical three-dimensional encoder-decoder operating on input patches measuring 160 by 160 by 160 voxels. The encoder reduces spatial resolution in five stages while increasing feature channels from 48 to 768. A bottleneck at 5 by 5 by 5 consolidates abstract features, and the decoder restores resolution through transposed convolutions, encoder skip connections and additional CFT-Triplet blocks. A scale-aware mixture-of-experts mechanism combines predictions from four decoder stages. Intermediate experts, with resolutions from 10 to 80 voxels per side, contribute differently to coarse shape and fine boundary detection. Their probabilities are resized, softmax-normalized and combined using instance-specific weights generated by global average pooling, a multilayer perceptron and softmax normalization. The training objective combined Tversky loss, which penalized false negatives more heavily, with weighted binary cross-entropy, which addressed the dominance of background voxels.

The retrospective dataset comprised 400 chest CT scans selected from 500 cases collected at Shandong Provincial Chest Hospital between June 2014 and November 2020. One hundred scans were excluded because of motion or metallic artifacts, incomplete lung coverage, unsuitable slice thickness or resolution, or severe pathology that obliterated the fissure anatomy. The remaining cases were divided at the patient level into 320 training scans, 40 validation scans and 40 test scans, a design intended to prevent data leakage. Testing therefore involved 40 held-out scans, not all 400 scans in the dataset. The scans came from Siemens, Philips and GE Medical Systems systems, with slice thicknesses of 0.625 to 1.25 millimetres and tube voltages of 100 to 120 kVp. Two experienced radiologists independently annotated fissures at voxel level, after which discrepancies were resolved in double-blind review by a senior specialist.

CFT-Triplet outperformed every comparator under a protocol that retrained all models with the same data splits, loss functions and hardware. Conventional CNN systems including U-Net, V-Net and Attention U-Net produced Dice scores of 78.85% to 81.15% and relatively high Betti number errors of 1.45 to 1.85, indicating more broken or fragmented predictions. Transformer models showed better continuity in the reported comparisons but were less precise at thin boundaries; nnFormer, for example, reached a Dice score of 83.92% and a normalized surface Dice of 86.85%. The Mamba-based U-Mamba and SegMamba-V2 performed better still. CFT-Triplet achieved 88.62% Dice, 88.18% precision, 92.15% normalized surface Dice at a one-millimetre tolerance and a Betti number error of 0.300. Its Dice improvement over SegMamba-V2 was associated with a Wilcoxon signed-rank p-value of 0.0083, while all reported comparisons met the study’s threshold of p below 0.05.

Ablation experiments supported the contribution and ordering of the model’s components. A texture-only encoder-decoder reached 80.45% Dice and a Betti number error of 1.650. Adding bidirectional Mamba2 raised Dice to 85.15% and reduced the topological error to 0.450. Arranging the Frequency Branch after contextual processing produced 87.50% Dice, 89.50% normalized surface Dice and a Betti number error of 0.350, outperforming parallel and frequency-first arrangements. Adding the multi-scale mixture of experts produced the full model’s 88.62% Dice and 0.300 Betti number error. Efficiency profiling at a 96-cubic-voxel input found 221.70 billion floating-point operations, 76 milliseconds of inference time and 9.4 gigabytes of GPU memory on one NVIDIA A100, compared with 240.20 billion operations, 135 milliseconds and 14.8 gigabytes for nnFormer and 328 billion operations, 112 milliseconds and 12.5 gigabytes for Swin UNETR.

The findings remain bounded by the study’s design. It was a retrospective evaluation on one private institutional dataset, and the model was tested on only 40 held-out cases, so performance in unselected hospitals, populations or scanning protocols is not established. The authors also reported that the system could struggle when severe emphysema had physically destroyed the fissure boundaries, causing its contextual branch to estimate separations that differed slightly from annotated pseudo-boundaries. The research therefore demonstrates an algorithmic advance rather than a clinical trial: it shows that a deliberately engineered combination of local texture, efficient global context and explicit frequency refinement improved quantitative fissure detection under the tested conditions. The findings do not establish improved patient selection, treatment planning or clinical outcomes.

Subject of Research: Retrospective evaluation of CFT-Triplet for three-dimensional pulmonary fissure segmentation in chest CT scans

Article Title: CFT-Triplet: integrating contextual, frequency and textural features for fine-grained pulmonary fissure detection

Article References: Xu, L., Jiang, Y., Tao, R., Li, N., Chen, R., Su, C., Li, G., & Tang, F. (2026). CFT-Triplet: integrating contextual, frequency and textural features for fine-grained pulmonary fissure detection. Frontiers in Artificial Intelligence, 9, Article 1817367. https://doi.org/10.3389/frai.2026.1817367

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1817367

Keywords: pulmonary fissures, chest CT, medical AI, CFT-Triplet, Mamba2, frequency-domain learning, image segmentation, topological continuity, COPD, integrating, contextual, frequency