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

cs.CVOct 24, 2026

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

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

Organizations: Department of Data Science and Artificial Intelligence, Indian Institute of Information Technology Dharwad, Ittigatti Road, Dharwad, 580009, Karnataka, India. · Department of Computer Science and Engineering, Indian Institute of Information Technology Dharwad, Ittigatti Road, Dharwad, 580009, Karnataka, India.

Abstract

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.

Figures & tables

Figure 1: 2D slices from the MSD dataset with the CT slice (left), the ground truth segmentation mask (middle) and the overlap of both (right).

Figure 2: 2D slices from the LIDC-IDRI dataset with poor illumination conditions.

Figure 3: (a) Initial block (b) Regular bottleneck module (c) Downsampling bottleneck module (d) Upsampling bottleneck module

Name	Type	Integration	Output size	Stage
initial			$16\times 128\times 128$	Initial
bottleneck1.0	downsampling		$64\times 64\times 64$	Stage 1
bottleneck1.1			$64\times 64\times 64$
bottleneck1.2			$64\times 64\times 64$
bottleneck1.3			$64\times 64\times 64$
bottleneck1.4			$64\times 64\times 64$

Table 1: Model architecture for an input image resolution of

1\times 256\times 256

Hyperparameter	Value
Image dimensions	$256\times 256$
Batch size	16
No. of epochs	250
Optimizer	Adam
Learning rate	$1\times 10^{-4}$
Output activation head	Sigmoid

Table 2: Selected hyperparameter values for model training.

Figure 4: A sample of 9 augmented images of a CT slice after passing through the data augmentation pipeline.

Model	DSC	Precision	Recall	IoU
UNet Ronneberger et al. (2015)	0.905826	0.926147	0.886377	0.827863
SegNet Badrinarayanan et al. (2017)	0.900529	0.912512	0.888857	0.819057
ENet Paszke et al. (2016)	0.843147	0.880315	0.808991	0.728829
ENet + Skip	0.845799	0.871146	0.821885	0.732800
Proposed model	0.90812	0.917452	0.898976	0.831703

Table 3: Evaluation metric scores comparison over different architectures on the LIDC-IDRI dataset with threshold of 0.5. 0 0 footnotetext: These model have been evaluated through experimentation.

Figure 5: Precision-Recall Curves with threshold overlay for different architectures.

Model	DSC
UNet with MobileNetV2 encoder Riaz et al. (2023)	0.8793
UNet with MobileNetV3 encoder Suji et al. (2024) 1 1 1 These model have been evaluated through experimentation	0.8824
UNet with EfficientNetB0 encoder Suji et al. (2024) 1 1 1 These model have been evaluated through experimentation	0.8503
nnU-Net Ferrante et al. (2022)	0.9
3D-Res2UNet Xiao et al. (2020)	0.9530
RRc-Unet 3D Le and Saut (2023)	0.8777

Table 4: Dice Score Coefficient (DSC) comparison with existing architectures.

Figure 6: Tumor slice with the ground truth mask overlapped (left), heatmap overlay of Grad-CAM (middle) and model segmentation mask prediction (right).

Model	No. of parameters	GMAC	inference time (sec)
UNet Ronneberger et al. (2015)	7,781,761	42.35	168.05
SegNet Badrinarayanan et al. (2017)	9,374,593	26.68	93.94
UNet with MobileNetV2 encoder Riaz et al. (2023)	12,366,977	8.12	27.82
UNet with MobileNetV3 encoder Suji et al. (2024)	9,370,561	7.17	22.59
UNet with EfficientNetB0 encoder Suji et al. (2024)	14,233,597	8.12	31.29
3D-Res2UNet Xiao et al. (2020)	6,767,889	229.66	623.18

Table 5: Number of parameters, computational efficiency and inference time on CPU to process a CT volume of dimensions

256\times 256\times 384

Medical Image Segmentation

Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.

See all 788 papers

Parameter-Efficient Fine-Tuning of Foundation Models for Liver Tumor Segmentation in CT
Sep 12, 2026Ramtin Mojtahedi, Mohammad Hamghalam, Jacob J. Peoples +2Medical Image Segmentation Parameter-Efficient Fine-Tuning
School of Computing, Queen’s University, Kingston, ON, Canada · Toronto General Research Institute, University Health Network, Toronto, ON, Canada · Department of Electrical Engineering, Qa.C., Islamic Azad University, Qazvin, Iran +3
Mammographic Lesion Segmentation with Lightweight Models: A Comparative Study
Apr 26, 2026Helder OliveiraBreast Cancer Efficient Neural Network Inference
aIndependent Researcher, GnosisX, Calgary, AB, Canada
MambaLiteUNet: Cross-Gated Adaptive Feature Fusion for Robust Skin Lesion Segmentation
Apr 22, 2026Md Maklachur Rahman, Soon Ki Jung, Tracy HammondMamba Multi-Scale Feature Fusion
Texas A&M University, College Station, TX, USA · Kyungpook National University, Daegu, South Korea

Sep 12, 2026J/K move · Enter open · S save

Parameter-Efficient Fine-Tuning of Foundation Models for Liver Tumor Segmentation in CT

Ramtin Mojtahedi, Mohammad Hamghalam, Jacob J. Peoples, Richard K. G. Do+1

School of Computing, Queen’s University, Kingston, ON, Canada · Toronto General Research Institute, University Health Network, Toronto, ON, Canada · Department of Electrical Engineering, Qa.C., Islamic Azad University, Qazvin, Iran+3

Medical Image Segmentation Parameter-Efficient Fine-Tuning Vision Foundation Model Adaptation Medical Imaging Foundation Models Promptable Image Segmentation

Open paper AskPDF arXiv

We evaluated parameter-efficient fine-tuning (PEFT) of the Segment Anything Model (SAM) for liver tumor segmentation in abdominal CT of colorectal liver metastases. We compared Low-Rank Adaptation (LoRA), 4-bit Quantized LoRA (QLoRA), a convolutional adapter (Conv-Adapter), and our Directional Spectral Top-K adapter (DiSCo), training only adapters while freezing the SAM backbone. DiSCo derives spectral bases from singular value decomposition of row-normalized weights and learns rank-gated spectral coefficients, per-output magnitude offsets, and a spectral gain, with optional Top-K rank selection at inference and 0.14 M trainable parameters. We benchmarked five prompting regimes: no prompt, single-point, multi-point, and bounding boxes at intersection over union 0.50 and 0.75. Conv-Adapter and LoRA achieved the highest accuracy (overall Dice 0.793 and 0.792; single-point Dice 0.795 and 0.792; 95th-percentile Hausdorff distance (HD95) 32 mm). QLoRA was close (overall Dice 0.766; single-point Dice 0.768; HD95 36.41 mm), with 0.91 M trainable parameters, 120 ms latency, and 4.9 GB peak memory. DiSCo achieved the highest Dice per million trainable parameters (4.66), with overall Dice 0.653, single-point Dice 0.698, and HD95 49.53 mm. These results show an accuracy-efficiency trade-off and support PEFT for liver tumor segmentation with reduced adaptation costs when compute and labeled data are limited. Code: https://github.com/Ramtin-Mojtahedi/PEFT-SAM-Liver-CT

Abstract

Medical Image Segmentation

Explore similar work

Parameter-Efficient Fine-Tuning of Foundation Models for Liver Tumor Segmentation in CT

Mammographic Lesion Segmentation with Lightweight Models: A Comparative Study

MambaLiteUNet: Cross-Gated Adaptive Feature Fusion for Robust Skin Lesion Segmentation