SPEEDNet: Salient Pyramidal Enhancement Encoder-Decoder Network for Colonoscopy Images

Tushir Sahu, Vidhi Bhatt, Sparsh Mittal, R. Sai Chandra Teja, Nagesh Kumar S

TL;DR

SPEEDNet is a novel, lightweight encoder-decoder architecture designed to precisely segment lesions in colonoscopy images. By introducing the Dilated-Involutional Pyramidal Convolution Fusion (DIPC) block, the model achieves state-of-the-art performance (0.952 Dice score) while maintaining a tiny footprint of just 9.81 MB—significantly smaller than traditional models like UNet or FeedNet.

What Problem Does This Paper Address?

Colon cancer is a leading cause of cancer-related deaths, yet early detection through colonoscopy can dramatically improve survival rates. The challenge lies in the visual complexity of these images: lesions can be tiny, have blurry boundaries, or look remarkably similar to healthy tissue.

Existing deep learning models often face a trade-off: they are either too computationally heavy for real-time clinical use (high parameter count) or they struggle to capture "multi-scale" information—meaning they might miss a small polyp while focusing on a large tissue structure. Furthermore, standard convolutions are "spatial-agnostic," applying the same weights across the whole image, which isn't always ideal for the highly variable textures found in medical imaging.

Methods

The authors propose SPEEDNet, which builds upon the classic UNet "U-shaped" architecture but replaces standard bottleneck components with a specialized DIPC Block.

Key Innovations:

  • Involution vs. Convolution: Unlike standard convolution, which is channel-specific but spatial-agnostic, involution is spatial-specific and channel-agnostic. This allows the network to generate kernels that adapt to specific spatial locations, reducing redundancy and significantly lowering the parameter count.
  • DIPC Block: This block uses a pyramidal structure with varying dilation rates (1, 2, 4, and 8). By merging these dilated layers, the model creates a "pyramidal" receptive field that can capture both fine, intricate edges and large, prominent structures simultaneously.
  • Attention-Driven Enhancement: The network incorporates "salient maps" derived from downsampled versions of the input image. These maps help the encoder focus on the most informative regions of the tissue.
  • Tversky Loss: To combat the class imbalance common in medical datasets (where healthy tissue far outweighs diseased tissue), the authors used the Tversky loss function, which allows for fine-tuning the penalty for false negatives versus false positives.

Datasets

The model was evaluated on the EBHI-Seg dataset, a comprehensive 6-class biopsy image dataset. It includes:

  • Normal tissue
  • Polyps
  • High-grade Intraepithelial Neoplasia (IN)
  • Low-grade IN
  • Adenocarcinoma
  • Serrated adenoma

The dataset contains 2,226 labeled images (224x224 resolution), split 80:20 for training and testing.

Key Insights and Ideas

  • Efficiency is King: The most striking "aha" moment is the parameter efficiency. SPEEDNet uses only 2.4 million parameters. For comparison, FeedNet uses roughly 77 times more memory (185.58 MB vs. 9.81 MB) but achieves lower accuracy.
  • Spatial Adaptivity: By using involution, the model acknowledges that a pixel's importance depends on its specific location in the colon, not just its channel intensity. This helps in "de-blurring" the resulting segmentation masks.
  • The Power of Dilation: Using a pyramid of dilations ensures the network doesn't "forget" the big picture while looking at the details, a common failure point in deep encoders.

Key Results

SPEEDNet outperformed established benchmarks (UNet, FeedNet, AttResDUNet) across nearly all metrics:

  • Dice Score: Achieved an average of 0.953, outperforming the next best model (UNet+TL) at 0.937.
  • Recall: Reached 0.971, indicating the model is exceptionally good at catching actual lesions (minimizing false negatives).
  • Model Size: At 9.81 MB, it is roughly 57% smaller than a standard UNet (22.84 MB) and a fraction of the size of AttResDUNet (140.09 MB).
  • Qualitative Quality: Visual results show that SPEEDNet produces much cleaner, more continuous masks, whereas other models often produce "patchy" or fragmented segmentations.

Limitations and Future Work

While SPEEDNet performs admirably, the authors note a few areas for growth:

  • Class Imbalance: For specific rare classes like Serrated Adenoma, the predicted masks still differ visibly from the ground truth. The authors suggest that adding more diverse images for these specific classes would improve generalization.
  • Temporal Data: The current model works on static images. The authors plan to explore fusing temporal sequences (video frames) to take advantage of the motion information in actual colonoscopy procedures.
  • Unsupervised Learning: Future iterations may look into using unsupervised or semi-supervised learning to leverage the vast amounts of unlabeled medical data available.

Summary updated

Abstract

Accurate identification and precise delineation of regions of significance, such as tumors or lesions, is a pivotal goal in medical imaging analysis. This paper proposes SPEEDNet, a novel architecture for precisely segmenting lesions within colonoscopy images. SPEEDNet uses a novel block named Dilated-Involutional Pyramidal Convolution Fusion (DIPC). A DIPC block combines the dilated involution layers pairwise into a pyramidal structure to convert the feature maps into a compact space. This lowers the total number of parameters while improving the learning of representations across an optimal receptive field, thereby reducing the blurring effect. On the EBHISeg dataset, SPEEDNet outperforms three previous networks: UNet, FeedNet, and AttesResDUNet. Specifically, SPEEDNet attains an average dice score of 0.952 and a recall of 0.971. Qualitative results and ablation studies provide additional insights into the effectiveness of SPEEDNet. The model size of SPEEDNet is 9.81 MB, significantly smaller than that of UNet (22.84 MB), FeedNet(185.58 MB), and AttesResDUNet (140.09 MB).
BibTeX
@inproceedings{sahu2024speednet,
  author = {Tushir Sahu and Vidhi Bhatt and Sparsh Mittal and R. Sai Chandra Teja and Nagesh Kumar S},
  title = {{SPEEDNet: Salient Pyramidal Enhancement Encoder-Decoder Network for Colonoscopy Images}},
  booktitle = {iSES},
  year = {2024},
  doi = {10.1109/ises63344.2024.00039},
}
Figure