Early polyp segmentation techniques primarily relied on low-level feature processing, such as texture, color, and geometric characteristics [1]. Methods to focus on these features include region growing, watershed, and active contour analysis. The method of Sasmal et al. [16] consists of principal component pursuit and active contour model. The region-based method [17] divides the image into multiple regions to judge the features within regions respectively. A commonly used morphology-based method [17] is to perform preprocessing and subsequent processing on images to enhance edge features. The polyp segmentation method proposed by Gross et al. [18] combines Canny operator, nonlinear diffusion filtering, and other methods. However, due to the high similarity between polyps and the surrounding tissue, these traditional methods often struggle with accuracy, leading to an increased likelihood of missed or incorrect detections.

CNN-based methods [19–23] have made significant contributions to the development of polyp segmentation and have outperformed traditional methods in feature modeling, noise suppression, inference speed, and generalization ability. Akbari et al. [24] proposed a polyp segmentation model using a fully convolutional neural network, which used a new image patch selection method in the training phase, and the probability map generated by the network was effectively post-processed in the test phase. Brandao et al. [25] obtained the shape through a shading strategy and used it to recover the depth, and used the RGB model to receive the result to enrich the feature representation. Encoder-decoder-based models have demonstrated impressive performance in image segmentation. UNet [1] aggregates the encoder features with the upsampled features of the decoder through skip connections to generate high-resolution segmentation maps. UNet++ [3] links the encoder and decoder through nested and dense skip connections. The skip connections of UNet 3+ [5] include full-size internal connections between decoder blocks. Dilated convolutions extract and aggregate high-level semantic features with resolution preservation to achieve improvements in the encoder network. With the advancement of computer vision, ResNet [4] has become a widely adopted backbone in medical image segmentation methods. In our proposed PVT-DMHFR, the PreActBottleneck of ResNet is employed to optimize feature perception in MHFRs. Mask R-CNN [26] was adapted with a deeper feature extractor [27] for polyp segmentation. PraNet [2] generated a global attention feature map by inverting attention and used it to derive boundary information. PNS-Net [28] incorporated temporal and spatial cues based on self-attention for video polyp segmentation, while Spatial-Temporal Feature Transformation [29] aggregated features from adjacent frames to achieve notable performance in video polyp segmentation. U-KAN [30] redesigned the UNet by integrating the dedicated Kolmogorov-Arnold Networks (KAN) layers on the tokenized intermediate representation to improve segmentation performance. GCN-DE [31] projects both support and query images into a feature space, computes long-range and short-range dependencies within a global correlation module that processes the embeddings to reduce complexity and applies discriminative regularization to draw features of the same foreground class closer together, enhancing the accuracy of few-shot medical image segmentation. Polyp-Net [32] utilized a local gradient weighting-embedded level-set method, effectively reducing false-positive instances caused by high-intensity regions during prediction. UACANet [33] modified the U-Net shape network, and in its prediction module, the foreground, background and uncertainty region maps are aggregated with features, and the saliency map that provides this calculation assistance is calculated and propagated to the next prediction module. SANet [34] minimized the impact of irrelevant features on predictions through color transformation. MSNet [35] reduced complementary and redundant information across multi-scale features via a multi-scale subtraction network. While these CNN-based networks have achieved good results, they share a common limitation: an inability to efficiently capture both global information and fine details simultaneously. The decoder struggles to effectively aggregate global features for enhanced information supplementation, the overall generalization performance is not satisfactory, and the results in polyp segmentation also require further improvement.

Transformer-based methods outperform CNN-based approaches due to their advantages in integrating global context, capturing long-range dependencies, and modeling features more effectively. DC-Net [36] is a dual context network that enhances segmentation performance by reshaping the original image and multiscale feature maps and integrating global and multiscale contextual information. Transfuse [10] combined CNN and transformer architectures through sequential and parallel connections in the encoder, performing well but limited in scalability due to the resulting increase in network size. Many methods [37–40] use PVTv2 as the encoder backbone. SSFormer [37] sought to simplify the structure of the transformer encoder’s backend, aiming to reduce model parameters while enhancing local information. However, this simplification led to a decline in performance. HSNet [38] combined CNN and Transformer in parallel within the decoder, though it did not fully address the differences in representation between the two architectures. Polyp-PVT [39] utilized the Cascaded Fusion Module (CFM) and Camouflage Identification Module (CIM) to merging features and directly extract intricate details from low-level features, respectively. Additionally, the Similarity Aggregation Module (SAM) was implemented to investigate higher-order relationships between the local features of low-level from CIM and the cues of high-level from CFM. Nevertheless, Polyp-PVT did not thoroughly explore and augment the encoder’s output information, resulting in suboptimal feature aggregation by CFM. Furthermore, CIM’s direct extraction of detailed information from low-level features introduced noise. PVT-CASCADE [40] accurately identified the most critical local features through its CASCADE module, which included an Attention Gate (AG), a multi-stage loss and feature aggregation component, as well as an upconv module. However, the upconv module risked losing fine-grained details, and the CASCADE module could result in redundant features, especially when dealing with objects with indistinct edges.

In conclusion, although previous methods have yielded impressive results in addressing the polyp segmentation challenge, they often fall short of effectively bridging the semantic gap between Transformer and CNN architectures. This unresolved gap can negatively affect network performance. Additionally, many Transformer-based polyp segmentation models struggle to adequately process the four pyramid features of varying granularities generated by the encoder backbone, which encapsulates rich spatial details and semantic information. To address these issues, we propose a novel decoder with multi-headed feature receptors.

The network architecture of PVT-DMHFR is illustrated in Fig. 1. It consists primarily of a Transformer encoder and DMHFR. DMHFR mainly consists of MHFRs, FcaNet, and axial transformers. PVTv2 [11] functions as the Transformer encoder to capture features that represent long-range dependencies across multiple scales from the input image. As shown in Fig. 1a,b, DMHFR receives four pyramid features from the PVTv2 encoder (X1, X2, X3, and X4), which are then modeled in the frequency domain by FcaNet [15] to effectively capture the details and texture information of the image. These features are subsequently divided into two sets of three-element features: {X4′,X3′,X2′}, {X3′,X2′,X1′}, as well as one set of four-element features: {X4′,X3′,X2′,X1′}, they are respectively coarse-grained feature group, fine-grained feature group, and full-set feature group. These feature sets are fed into THFR32, THFR64, and FHFR128, which receive groups of features with the channel unified as 32, 64, and 128 respectively, then output after passing through the axial transformer [41].

DMHFR allows features of multiple different dimensional combinations to be perceived in parallel and be fused, retaining the features in the input image to a high degree, and most of the context output by MHFRs can be calculated in parallel by the axial transformer to express the global dependencies of features. This architecture achieves SOTA performance on several polyp segmentation benchmarks. Details are presented in the experimental section.

Recent studies [8,13,14,42] on vision tasks have demonstrated that transformer-based pyramid structures are better than CNNs in terms of generalization, robustness, and capturing multi-scale and multi-level features. In this proposed method, PVTv2 [11] is employed to extract multi-scale features. Unlike traditional Transformers, PVTv2 does not use a patch embedding module but uses convolution operations to consistently capture spatial information, delivering state-of-the-art performance across various dense prediction applications. PVTv2 generates pyramid features Xi∈{(88,88,64),(44,44,128),(22,22,320),(11,11,512)} according to input image I∈RH×W×3. These features are subsequently fed into the DMHFR.

Due to the high similarity between polyps and backgrounds and the limited ability of existing Transformer-based models to process (local) contextual information among pixels, the localization of local features with higher discrimination in the segmentation task is challenging. To address this challenge, we propose a novel Decoder with Multi-Head Feature Receptors (DHMFR) for pyramid features.

As Fig. 1b shows, DHMFR includes FcaNet [15] to model features in the frequency domain, perceive details and texture information of images, our proposed MHFRs (two THFRs and one FHFR) to perceive and fuse pyramid features in parallel, and axial transformer [41] to keep the full expressiveness of joint distribution over features. The features of four different dimensions from the encoder backbone are passed through four FcaNet blocks, and then the four features are grouped into two three-element feature groups and one four-element feature group, these three feature groups are coarse-grained feature group, a fine-grained feature group, and full-set feature group. The channel of each feature in feature groups: {X4′,X3′,X2′}, {X3′,X2′,X1′}, and {X4′,X3′,X2′,X1′} are adjusted to 32, 64, and 128, respectively, and they are passed to THFR32, THFR64, and FHFR128 accordingly. In MHFRs, the features within each group are perceived and fused in parallel, generating three output features. Next, these features are processed in parallel by the axial transformer, allowing the model to capture the global dependencies among features while preserving the full expressiveness of their joint distribution. Then, the channel of the output features will be adjusted to 32. Notably, the feature processed by FHFR128 is copied as an auxiliary prediction. This is because FHFR128 perceives and fuses all pyramid features in parallel, so its output is assumed to have a higher priority and designed to carry greater weight in the prediction output. Finally, the three output features P1,P2,P3, and an auxiliary output feature P3_aux are sent to the prediction head, and the four different predictions are fused to generate the final segmentation map.

We utilize additive aggregation with four prediction heads to compute the final prediction map, as expressed in Eq. (5): (5)output=pw1×P1+pw2×P2+pw3×P3+pw4×P3_auxwhere P1, P2, P3, and P3_aux represent feature maps from four prediction heads, and pw1, pw2, pw3 and pw4 denote the weights assigned to each feature map in the final prediction map. In PVT-DMHFR, pwi=1.0,i∈{1,2,3,4}.

The loss function for feature maps of each prediction head can be expressed as Eq. (6): (6)lossp=LwIoU(P,G)+LwBCE(P,G)where lossP refers to each loss for prediction heads, LwIoU(⋅) denotes weight intersection over union (IoU) loss, and LwBCE(⋅) denotes weight binary cross entropy (BCE) loss, this combination of loss functions imposes restrictions on the prediction map regarding local details (pixel level) and global structure (object level). The final loss for each prediction head is separately computed and then aggregated as Eq. (7): (7)loss=lw1×lossP1+lw2×lossP2+lw3×loss2+lw4×lossP3_auxwhere lwi=1.0,i∈{1,2,3,4}.

We validate the performance of the proposed method on five public polyp datasets: CVC-ClinicDB [44] has 612 polyp images extracted from 31 colonoscopy videos. Kvasir-SEG [45] has 1000 polyp images collected from the Kvasir dataset’s polyp class. CVC-T [46], a subset of the EndoScene dataset, has 60 polyp images. CVC-ColonDB [47] has 380 polyp images. ETIS-LaribPolypDB [48] has 196 polyp images.

We utilize several key metrics as PraNet [2] used to assess the performance of our proposed method comprehensively. Mean Dice (mDic) [49] quantifies the overlap between predicted and ground truth segmentations, producing a score between 0 and 1, with higher values indicating better accuracy. Mean intersection over union (mIoU) also measures overlap but is more stringent, providing a ratio of the intersection to the union of the predicted and actual regions. Mean absolute error (MAE) calculates the average absolute differences between predictions and ground truth, providing insight into overall accuracy. Weighted F-measure (Fβω) [50] balances precision and recall, particularly useful in scenarios with class imbalance. S-measure (Sα) [51] evaluates structural similarity by considering region and boundary adherence, while E-measure (Eξ) [52,53] extends this by incorporating boundary precision and region consistency, we report the both mean value of E-measure (mEξ) and max value of E-measure (maxEξ). The equation of mDic is given in Eq. (8), and the equation of mIoU is given in Eq. (9): (8)mDic(P,G)=2×|A⋂B||A|+|B|=2×TP2×TP+FP+FN(9)mIoU(P,G)=A⋂BA⋃B=TPTP+FP+FNwhere P refers to prediction map, G refers to the Ground Truth, TP denotes true positive instances, FP denotes false positive instances, FN denotes false negative instances. The equation of Fβω is given in Eq. (10):

(10){R=TPTP+FNP=TPTP+FPFβω=(1+β2)Pω⋅Rωβ2⋅Pω+Rωwhere R refers to recall, P refers to precision, β is a parameter to trade-off R and P. The MAE score is computed by Eq. (11):

(11)MAE=1H×W∑x=1H∑y=1W|P(x,y)−G(x,y)|where H refers to height of images, W refers to width of images. The Sα is computed by Eq. (12):

(12)Sα=αS0+(1−α)Srwhere Sr and S0 refers to the region-aware and object-aware similarity measure, and the trade-off coefficient, α, is set to 50 by default. The equation of Eξ is given in Eq. (13):

(13)Eξ=1H×W∑x=1H∑y=1Wϕ(x,y)where ϕ(x,y) denotes the enhanced alignment matrix that capture pixel-level matching and image-level statistics.

Our proposed method is implemented using the PyTorch 2.0.0 framework. The model is trained using an NVIDIA RTX 3090 GPU with 24 GB of memory. We use the Adam optimizer [54] and set the learning rate to 5 × 10⁻⁵ without decay. Following Polyp-PVT [39], we use a multi-scale {0.75, 1.0, 1.25} training strategy with gradient clipping set to 0.5, configure the batch size to 16, set the maximum number of epochs to 100, and resize input images to 352 × 352 pixels. To ensure fairness in our comparative experiments, we adopt the same data division method as used in PraNet. A total of 900 images from Kvasir-SEG and 548 images from CVC-ClinicDB are used as training sets, while the remaining 100 images from Kvasir-SEG and 64 images from CVC-ClinicDB are reserved as test sets for evaluating the model’s learning ability. We further assess the model’s generalization ability on three additional datasets: CVC-ColonDB, CVC-T, and ETIS-Larib. In this paper, we compare the proposed method with 15 state-of-the-art image segmentation models, including UNet [1], UNet++ [3], PraNet [2], MSNet [35], SANet [34], Transfuse [10], UACANet [33], TMF-Net [55], C2F-Net [21], Polyp-PVT [39], SSFormer [37], ColonFormer [56], ESFPNet [57], FCBFormer [58], and PVT-CASCADE [40].

Table 1 presents the quantitative results of the feature modeling capabilities comparison between PVT-DMHFR and 15 different SOTA methods trained on the ClinicDB and Kvasir-SEG datasets. The PVT-DMHFR exhibits superior feature modeling performance compared to other methods. Compared to the transformer-based method PVT-Cascade, our approach improves the mDic and mIoU on the CVC-ClinicDB dataset by 1.2% and 2.1%, respectively. Additionally, it increases the Fβω by 1.4%, enhances the Sα by 1.2%, improves the mEξ by 1.3%, improves the maxEξ by 1.5%, and reduces the MAE by 0.6%.

On the Kvasir-SEG dataset, compared to the best CNN-based method, UACANet, our method achieves a 0.6% improvement in the mDic, a 0.4% increase in mIoU, a 1.3% rise in the Fβω, a 1.0% gain in the Sα, a 0.9% boost in the mEξ, improves the maxEξ by 0.2%, and a 0.4% reduction in the MAE. In conclusion, our approach delivers top performance across most metrics on the Kvasir-SEG dataset, with the exception of the maxEξ. On the ClinicDB dataset, our method outperforms all others, achieving the highest scores across all evaluation metrics.

Tables 2 and 3 present the results of performance comparison between our PVT-DMHFR and 15 methods on three unseen datasets: CVC-T, CVC-ColonDB, and ETIS-Larib. On the CVC-T dataset, our method achieves an mDic score of 0.8% higher than the best CNN-based method, SANet, and 0.7% higher than the best transformer-based method, PVT-Cascade. Our method performs best on all metrics, except for maxEξ and MAE, where it lags ColonFormer by 0.9% in maxEξ and SSFormer by 0.1% in MAE. On the CVC-ColonDB dataset, our method surpasses the best CNN-based method, SANet, by 5.7% in mDic, and the transformer-based method, PVT-Cascade, by 0.9%. Our method performs best on all metrics except Sα and MAE, which lag behind Polyp-PVT by 0.2% and 0.3%. On the ETISLarib dataset, our method achieves a mDic score of 4.4% higher than the best CNN-based method SANet and 0.6% higher than the best transformer-based method PVT-Cascade. Our method performs best on all metrics, except for MAE, which lags behind PVT-Cascade by 0.3%.

Table 4 presents the results of performance of combination of DMHFR and different backbones, including hierarchical transformer-based backbones (PVTv2b1 [11], PVTv2b2 [11], PVTv2b3 [11], PVTv2b4 [11], PVTv2b5 [11], mitb5 [14]), transformer-based backbone (R50+ViT-B_16 [8]), and CNN-based backbone (ResNetV2 [43]). The results demonstrate that hierarchical transformer-based backbones, particularly PVTv2b3 and PVTv2b2, consistently outperform both transformer-based (R50+ViT-B_16) and CNN-based (ResNetV2) backbones across all datasets and metrics.

For example, PVTv2b3 achieves the highest mDic (0.943) and mIoU (0.899) on the CVC-ClinicDB dataset, while PVTv2b5 excels on Kvasir-SEG with mDic of 0.922 and mIoU of 0.876. Additionally, mitb5 demonstrates well-balanced and strong performance across all five datasets. In contrast, the transformer-based backbone typically performs significantly worse than its hierarchical counterparts, achieving top scores only in one case (an mDic of 0.931 on CVC-ClinicDB). Meanwhile, the CNN-based backbone consistently underperforms, underscoring their limited effectiveness in these tasks when working in conjunction with DMHFR. This performance disparity can be attributed to the alignment between the hierarchical transformer-based backbones and DMHFR’s input requirements. The hierarchical transformer-based backbones naturally produce outputs with dimensions that align with DMHFR’s architecture, minimizing the need for extensive preprocessing or additional layers. Conversely, both transformer-based and CNN-based backbones require supplementary adjustments, such as additional transformations or layers, to ensure compatibility with DMHFR. These modifications not only introduce computational overhead but may also disrupt the original feature distributions, leading to reduced performance.

By contrast, the seamless integration of hierarchical transformer-based backbones with DMHFR enhances efficiency and preserves the integrity of feature representations. This synergy allows for streamlined processing and optimized information flow, resulting in superior performance across diverse datasets.

Overall, the combination of DMHFR and hierarchical transformer-based backbones demonstrates remarkable adaptability and effectiveness across diverse datasets.

Based on the above analysis, our PVT-DMHFR shows impressive learning and generalization capabilities on the challenging task of polyp segmentation, as well as performance that is superior to other SOTA methods.

To thoroughly evaluate the performance of our proposed method, we compared our PVT-DMHFR with SOTA methods in terms of visual results. As shown in Fig. 4, our PVT-DMHFR demonstrates several advantages over these SOTA methods: First, the transformer-based encoder backbone we employed, PVTv2, enhances polyp localization accuracy. Second, PVT-DMHFR consistently produces segmentation results with high accuracy for polyps of various sizes and shapes. This stability and accuracy are largely attributed to the proposed MHFRs, which effectively capture and fuse multiple groups of multi-scale information. Additionally, the integration of FcaNet and the axial transformer, applied before and after the MHFRs, strengthens the model’s ability to extract features from the encoder backbone and capture long-range dependencies within the feature map, significantly improving overall feature representation.

We comprehensively evaluate PVT-DMHFR and SOTA methods in terms of floating-point operations (FLOPs) and the number of parameters (Params). As shown in Table 5, the proposed PVT-DMHFR demonstrates an advantage over several strong competitors (e.g., PVT-Cascade, ColonFormer, SSFormer, FCBFormer, and ESFPNet) in terms of Params. However, it is slightly less efficient in FLOPs, surpassing only FCBFormer in this aspect. Overall, PVT-DMHFR achieves a well-balanced trade-off between computational efficiency and segmentation accuracy.

We performed ablation experiments to assess the contribution of each component in our method. The mDic and mIoU metrics were selected to represent network performance, and the results are summarized in Table 6. When the FcaNet modules were removed, the mDic and mIoU scores on the CVC-ClinicDB dataset dropped by 0.4% and 0.5%, respectively. This indicates that passing features through the FcaNet module before processing them with the MHFRs helps the model learn and represent features more effectively. On the ETIS-LaribPolypDB dataset, the mDic and mIoU scores decreased by 0.3% and 1.1%, respectively, suggesting that FcaNet improves the model’s generalization ability through superior feature capture. After removing the MHFRs, the mDic and mIoU scores fell by 1.1% and 0.8%, respectively, on the CVC-ClinicDB dataset, and by 1.4% and 1.6% on the CVC-ColonDB dataset. These results demonstrate that MHFRs play a crucial role in fusing multiple feature sets at various resolutions, leading to a significant improvement in segmentation accuracy. When the axial transformers were excluded, the mDic and mIoU scores on the CVC-ClinicDB dataset dropped by 0.5% and 1.6%, respectively. The mDic and mIoU scores on the CVC-ColonDB dataset dropped by 0.8% and 0.7%. This underscores the importance of axial transformers in capturing long-range dependencies within feature maps post-MHFR processing, further enhancing feature representation and improving both segmentation accuracy and generalization.

We present visualization results to better demonstrate the impact of our proposed MHFRs and integration of FcaNet modules and axial transformers in PVT-DMHFR. As illustrated in Fig. 5, the removal of any module from PVT-DMHFR results in a noticeable decline in segmentation accuracy. This performance degradation may stem from several factors: excluding FcaNet reduces the model’s capacity to capture detailed features, removing the axial transformers diminishes its ability to account for long-range feature dependencies, and omitting MHFRs impairs the fusion of multi-level features from the encoder backbone, leading to the loss of crucial semantic information.

Additionally, we explore the impact of replacing the Axial-Transformer with four alternative widely-used attention mechanisms: SE-Attention [62], CoT-Attention [59], EMA [60], and PSA [61], to evaluate their performance. As shown in Table 7, none of these attention mechanisms outperforms the Axial-Transformer. For instance, on CVC-ClinicDB, Axial-Transformer achieves the highest mDic (0.938) and mIoU (0.892), while PSA and CoT-Attention fall slightly short with mDic values of 0.936 and 0.932, and mIoU values of 0.889 and 0.887, respectively. Similarly, on the CVC-T dataset, the Axial-Transformer maintains its lead with mDic and mIoU scores of 0.897 and 0.833, respectively. While EMA and PSA achieve marginal improvements in mIoU on Kvasir-SEG and ETIS-LaribPolypDB, these gains are isolated and do not match the overall performance of the Axial-Transformer.

The consistent underperformance of SE-Attention and CoT-Attention reflects their limited ability to model global dependencies. In addition, while EMA and PSA perform better, their results lack consistency across datasets. Axial-Transformer’s superior ability to model long-range dependencies and spatial features explains its dominance, making it the most effective attention mechanism in this study.