PaperlayerView exact revision on arXiv

Source-linked AI summary

Energy Efficient Lung Tumor Segmentation with a modified CBAM ENet Skip Model

Sharan Thummagunti, Aayushi Padia, Abhiram Koppuravuri, Prabhu Prasad B M, G N Girish

arXiv:2610.33479v1cs.CVcs.AIcs.LG

TL;DR

Computationally expensive deep-learning segmentation can require GPU acceleration for real-time inference, limiting deployment in resource-constrained clinical settings. The model combines ENet with CBAM attention modules and UNet-style encoder–decoder skip connections. Recall 0.898976 and IoU 0.831703 with a threshold of 0.5; up to 10× faster, requiring 79× fewer GMAC and having 22× less parameters.

  • Problem

    Computationally expensive deep-learning segmentation can require GPU acceleration for real-time inference, limiting deployment in resource-constrained clinical settings.

    Source: §1 · p. 2

  • Method

    The model combines ENet with CBAM attention modules and UNet-style encoder–decoder skip connections.

    Source: §1 · p. 3

  • Results

    Recall 0.898976 and IoU 0.831703 with a threshold of 0.5; up to 10× faster, requiring 79× fewer GMAC and having 22× less parameters.

    Source: §5 · p. 18

  • Takeaways & Limitations

    The reported efficiency and speed are presented as making the model suited to resource-constrained settings, including mobile health units and rural clinical settings.

    Source: §1 · pp. 3–4

    Grad-CAM indicates difficulty identifying thin tumor branches away from the main tumor volume with high confidence; 2D slices also lose 3D context.

    Source: §4.3 · p. 17

Abstract

from arXiv · show

Lung tumor segmentation plays a vital role in early detection and treatment planning of lung cancer. Although state-of-the-art deep learning models achieve high dice score coefficients (DSC), each forward pass requires computing large amounts of floating point operations, making them computationally expensive, requiring GPU acceleration, and having slow inference speed, limiting real-world clinical deployment. We propose a lightweight, energy-efficient model that leverages the ENet architecture for its real-time inference capability and low computational cost, enhanced with UNet-style skip connections between encoder and decoder stages for feature reuse and the lightweight Convolutional Block Attention Module (CBAM) for improved feature representation. The model is trained on publicly available data from The Cancer Imaging Archive (TCIA) via the Medical Segmentation Decathlon Challenge (MSD) and evaluated on the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset. Our model achieves a DSC of 0.90812 (threshold of 0.5) with its compact architecture of 430,052 parameters and 0.549 Giga Multiply-Accumulate operations (GMAC), making it up to 10 times faster, requires 79 times less GMAC than UNet, and has 22 times fewer parameters than SegNet. These characteristics make it well-suited for deployment in clinical settings where computational resources are limited and decrease the cost of utilizing such a system. Additionally, the explainable AI technique of Grad-CAM is used to provide visual interpretability of the model's predictions and to identify regions where segmentation performance requires improvement.

1 Introduction

High computational costs and GPU requirements limit deep-learning lung tumor segmentation in resource-constrained settings. The paper proposes an ENet model augmented with CBAM and UNet-style skip connections to balance segmentation performance and computational efficiency.

1 Introduction

  • Deep models' costly 3D convolutions can require GPU acceleration for real-time inference, constraining deployment in small clinics and mobile health units.Source: §1 · p. 2
  • UNet and SegNet's VGG-16 base and transformer modules raise GMAC and inference time, limiting real-time use in resource-constrained settings.Source: §1 · p. 2
  • CBAM applies channel and spatial attention sequentially, using average and max pooling for channel attention with minimal computational overhead.The channel module identifies what is important, while the spatial module focuses on where informative features lie.Source: §1 · p. 3
  • ENet offers a speed–accuracy balance; adding UNet-style skips reuses features and mitigates the need for encoder pretraining.The paper proposes CBAM alongside the skip connections to enhance ENet's representation capabilities.Source: §1 · p. 3
  • Up to 10× faster and 79× fewer GMAC than UNet, with 22× fewer parameters than SegNet, the model targets resource-constrained clinical settings.The introduction also describes rural clinical settings and reduced usage costs as intended benefits.Source: §1 · pp. 3–4

2 Related Work

Related work covers 2D CNNs, 3D CNNs, and transformer-based approaches to medical segmentation. These approaches trade feature representation or volumetric context against computational and memory demands.

2 Related Work

  • Prior lung-tumor and nodule CNN research is grouped into 2D convolutional, 3D-based, and transformer-based methods.Source: §2 · p. 4

2.1 2D convolutional neural network

  • 2D U-Net uses encoder–decoder skip connections for high-resolution feature fusion, while slice-wise processing can lose inter-slice continuity and 3D context.SegNet instead uses max-unpooling; it requires 1.6× less GMAC than UNet, according to the passage.Source: §2.1 · p. 4

2.2 3D convolutional neural network

  • 3D U-Net learns volumetric context useful for delineating subtle nodules across adjacent slices, but its memory footprint and training time pose practical limitations.The related work describes 3D models as performing better than 2D counterparts, especially for small or complex lesions.Source: §2.2 · pp. 4–5

2.3 Transformer architectures

  • UNETR uses a ViT encoder and volumetric patches; Swin-UNETR uses shifted-window attention, but transformer modules have high memory demands and slower inference than CNNs.Source: §2.3 · p. 5

3 Methods

The methods adapt ENet for 2D lung-tumor segmentation, combining encoder–decoder skip connections and CBAM with dataset-specific preprocessing and augmentation. Training uses MSD data and evaluation uses LIDC-IDRI scans.

3.1 Dataset

  • MSD Task 6 supplies lung-tumor CT scans, converted into 2D slices; slices without tumors outnumber tumor-containing slices by roughly 8.5 to 1.Tumor-slice oversampling is used to address this imbalance.Source: §3.1 · p. 6
  • Testing uses expert-annotated LIDC-IDRI volumes, whose image quality and intensity distributions differ from MSD; both datasets' intensities are clipped to −1024 to 3071 HU.LIDC-IDRI annotations come from a two-phase process involving four thoracic radiologists.Source: §3.1 · pp. 6–7

3.2 Network Architecture

  • Modified ENet uses early downsampling and a deep, narrow design, with CBAM in regular encoder bottlenecks and skip connections carrying encoder features to the decoder.The skip connections link stage 1 and the initial block to decoder stages.Source: §3.2 · pp. 7, 8, 10
  • Adding skip connections removed the initial training plateau and reduced convergence time by 32%; extra attention gates did not improve performance and added computation.Without skips, training plateaued for 100 epochs before a performance jump.Source: §3.2 · p. 10

3.3 Model Training

  • Binary cross-entropy allows flexible probability outputs, and applying a threshold produced metric-score increases of up to 6.5%, unlike dice-loss training.Training ran for 250 epochs with 1 × 256 × 256 inputs and batch size 16.Source: §3.3 · pp. 10–11

3.4 Data Preparation and Augmentation

  • Preprocessing aligns scans to left-anterior-superior orientation, clips and standardizes intensities, and skips the first 30 axial slices.Training with a lung window of 800 to −1024 degraded segmentation performance.Source: §3.4 · p. 11
  • The augmentation pipeline applies horizontal and vertical flips, linear contrast, Gaussian noise and blur, affine transformations, and elastic deformation.Experiments found Gaussian noise and blur helped reduce overfitting; the authors also report that 2D slices allow the chosen flips and rotations.Source: §3.4 · pp. 11–12

4 Results

On LIDC-IDRI, the proposed model achieved a DSC of 0.90812 at threshold 0.5, with strong precision-recall performance and better DSC than the leading architectures except 3D-Res2UNet. It remains faster and lighter than the other architectures considered, while Grad-CAM exposes difficulty capturing thin tumor branches.

4.1 Evaluation Metrics

  • At threshold 0.5, the proposed model reached DSC 0.90812, recall 0.898976, and IoU 0.831703; UNet led precision at 0.926147 versus 0.917452.The proposed model's DSC exceeded SegNet (0.900529), ENet with skip connections (0.845799), and baseline ENet (0.843147).Source: Table 3 · p. 13, §4.1 · pp. 13–14
  • The paired Wilcoxon test across LIDC-IDRI volumes led the authors to report a statistically significant, consistent DSC improvement over UNet despite a modest absolute gain.Source: §4.1 · p. 14

4.2 Model Performance

  • The proposed model achieved the highest AP (0.9738), ahead of UNet (0.9680), SegNet (0.9662), ENet with skip connections (0.9164), and baseline ENet (0.9070).It maintained high precision and recall at threshold 0.8.Source: Figure 5 · p. 14, §4.2 · pp. 14–15
  • The proposed model exceeded the DSC of leading architectures except 3D-Res2UNet and maintained DSC >0.90 across MSD and LIDC-IDRI without retraining.The comparison notes that 3D-Res2UNet used a larger LUNA16 training dataset.Source: §4.2 · p. 15
  • Grad-CAM shows focus on the main tumor but difficulty capturing thin outgrowths; at threshold 0.8, false negatives increase and recall decreases.The authors identify sensitivity to subtle tumor structures as a target for future refinement.Source: §4.2 · pp. 15–16, Figure 6 · p. 16
  • Despite added GMAC and inference time from skip connections and CBAM, the model remained faster and lighter than all considered architectures, without requiring hardware acceleration for real-time inference.Inference times were measured on an Intel Core i7-1165G7 CPU for a 256×256×384 CT scan.Source: Table 5 · p. 16, §4.2 · pp. 16–17

4.3 Discussions

  • The compact design meets strict GMAC constraints for resource-limited CAD without an accelerator, but limits representational capacity versus larger transformer models; 2D slices also lose 3D context.The discussion proposes 2.5D training with adjacent slices as a way to add 3D information while retaining low computational costs.Source: §4.3 · p. 17

5 Conclusion

The modified ENet combines CBAM attention with UNet-style skip connections for lung tumor segmentation. On LIDC-IDRI, it reports segmentation scores and computational-efficiency comparisons, while Grad-CAM highlights difficulty identifying thin tumor branches away from the main volume.

  • The architecture modifies ENet with lightweight CBAM attention modules and UNet-style encoder-decoder skip connections.The enhancements strengthen feature representation and attention capabilities with minimal computational overhead.Source: §5 · p. 17
  • At threshold 0.5, the LIDC-IDRI model achieves DSC 0.90812, precision 0.917452, recall 0.898976, and IoU 0.831703, outperforming UNet and SegNet.Source: §5 · pp. 17–18
  • The model is up to 10× faster, requires 79× fewer GMAC, and has 22× less parameters.Source: §5 · p. 18
  • Grad-CAM highlights difficulty identifying thin branches of tumor growth away from the main tumor volume with high confidence.Source: §5 · p. 18

Declarations

The declarations report no funding or conflicts of interest, state that no datasets were generated or analyzed, and assign the listed research and writing contributions to the authors.

Funding

  • The authors report receiving no funding support from any organization for the submitted work.Source: Funding · p. 18

Conflicts of Interests

Data Availability Statement

CRediT authorship contribution statement

Loading 2610.33479v1…