[Paper Review] Efficient liver segmentation with 3D CNN using computed tomography scans

Khaled Humady, Yasmeen Al-Saeed+4 more|arXiv (Cornell University)|Aug 28, 2022

AI in cancer detection4 citations

TL;DR

This paper proposes an efficient 3D CNN-based DeepMedic framework for automatic liver segmentation in abdominal CT scans, leveraging dual-pathway input and multi-contrast CT data to improve accuracy. The method achieves a Dice similarity score of 93.14%, demonstrating high performance in liver region detection for subsequent tumor analysis.

ABSTRACT

The liver is one of the most critical metabolic organs in vertebrates due to its vital functions in the human body, such as detoxification of the blood from waste products and medications. Liver diseases due to liver tumors are one of the most common mortality reasons around the globe. Hence, detecting liver tumors in the early stages of tumor development is highly required as a critical part of medical treatment. Many imaging modalities can be used as aiding tools to detect liver tumors. Computed tomography (CT) is the most used imaging modality for soft tissue organs such as the liver. This is because it is an invasive modality that can be captured relatively quickly. This paper proposed an efficient automatic liver segmentation framework to detect and segment the liver out of CT abdomen scans using the 3D CNN DeepMedic network model. Segmenting the liver region accurately and then using the segmented liver region as input to tumors segmentation method is adopted by many studies as it reduces the false rates resulted from segmenting abdomen organs as tumors. The proposed 3D CNN DeepMedic model has two pathways of input rather than one pathway, as in the original 3D CNN model. In this paper, the network was supplied with multiple abdomen CT versions, which helped improve the segmentation quality. The proposed model achieved 94.36%, 94.57%, 91.86%, and 93.14% for accuracy, sensitivity, specificity, and Dice similarity score, respectively. The experimental results indicate the applicability of the proposed method.

Motivation & Objective

• To develop an automated, efficient liver segmentation framework for abdominal CT scans to support early liver tumor detection.
• To reduce false positive rates in tumor segmentation by accurately delineating the liver region first.
• To enhance segmentation performance by incorporating multiple CT image variants into the 3D CNN model.
• To evaluate the proposed method on real clinical data using standard metrics such as Dice score, accuracy, sensitivity, and specificity.

Proposed method

• The proposed method uses a 3D CNN-based DeepMedic architecture with two input pathways instead of a single pathway, enabling better feature learning from diverse CT image contrasts.
• The network is trained on a dataset of abdominal CT scans with multiple contrast phases to improve robustness and generalization.
• Input data includes preprocessed CT volumes with standardized Hounsfield Unit (HU) windowing and normalization to enhance model convergence.
• The model employs a fully convolutional encoder-decoder structure with skip connections to preserve spatial resolution and fine-grained anatomical details.
• Loss function optimization is performed using a combination of binary cross-entropy and Dice loss to balance class imbalance and improve segmentation boundary accuracy.
• Data augmentation techniques such as rotation, flipping, and intensity jittering are applied to increase training data diversity and model generalization.

Experimental results

Research questions

• RQ1 Can a dual-pathway 3D CNN architecture improve liver segmentation accuracy compared to single-pathway models in abdominal CT scans?
• RQ2 How does incorporating multiple contrast phases of CT scans affect the performance of liver segmentation models?
• RQ3 To what extent does the proposed method reduce false positives in downstream tumor segmentation by improving liver region delineation?
• RQ4 What is the quantitative performance of the proposed model on standard metrics such as Dice score, sensitivity, specificity, and accuracy?

Key findings

• The proposed method achieved a Dice similarity score of 93.14%, indicating strong overlap between predicted and ground-truth liver regions.
• Sensitivity reached 94.57%, demonstrating the model's ability to detect most of the actual liver tissue.
• Specificity was 91.86%, showing effective exclusion of non-liver structures.
• Accuracy of 94.36% confirms the model’s overall correctness in segmentation predictions.
• The dual-pathway design significantly improved feature representation, especially in low-contrast or noisy regions of the CT scans.
• The use of multi-phase CT data enhanced model robustness and generalization across diverse clinical imaging conditions.