CRLM-GAN: a feature-constrained GAN-based deep learning framework for multi-parametric MRI-based segmentation of colorectal liver metastases before and after chemotherapy.

backgroundThe manual segmentation of colorectal liver metastases (CRLMs) is time-consuming and labour-intensive because of their high degree of heterogeneity, and identifying indistinct boundaries in post-treatment cases is more challenging. Automated segmentation techniques based on deep learning (DL) can alleviate these challenges. Generative adversarial networks (GANs) have emerged as an important development in deep learning for medical image segmentation. The aim of this study was to develop a GAN-based model for multi-parametric MRI-based CRLM segmentation and validate its clinical efficacy.

methodsThis study included a total of 641 CRLM cases who underwent multi-parametric magnetic resonance imaging (MRI) or contrast-enhanced computed tomography (CT). A retrospective cohort of 111 patients (444 cases under four conditions) with pathologically confirmed CRLMs was enrolled, and 2,546 two-dimensional tumour images were obtained. All patients underwent pre- and post-neoadjuvant chemotherapy (NACT) MRI scanning, including diffusion-weighted imaging (DWI) and T2-weighted imaging (T2WI). The dataset was split at a 6:4 ratio for training and testing. A GAN-based DL framework was proposed for CRLM segmentation, and five single-condition evaluations and four cross-sequence evaluations were systematically performed. The multi-feature constrained GAN-based model incorporated UNet++ as a generator and pre-trained ResNet-50 as a discriminator. The Dice similarity coefficient (DSC) was used as the primary evaluation metric. By extracting and fusing deep convolutional features, this approach utilized a multi-scale constrained strategy in the discriminator and combined binary cross-entropy with Dice loss in the generator. In addition, 197 publicly available contrast-enhanced CT scans containing 3,593 two-dimensional tumour images were collected to evaluate the adaptability of the model as a complementary modality.

resultsIn the single-condition evaluations, CRLM-GAN achieved DSCs of 0.81 (95% CI: 0.76, 0.85), 0.70 (95% CI: 0.63, 0.77), 0.67 (95% CI: 0.55, 0.77), and 0.61 (95% CI: 0.53, 0.68) on pre-NACT DWI, post-NACT DWI, pre-NACT T2WI, and post-NACT T2WI, respectively. Moreover, the model obtained a DSC of 0.70 (95% CI: 0.66, 0.73) on the public CT dataset. The results of cross-sequence experiments revealed that training the model with combined pre-/post-NACT DWI and T2WI data led to improvements in DSCs for both DWI and T2WI sequences.

conclusionsCRLM-GAN demonstrated superior one-stage segmentation performance across the multi-parametric MRI-based dataset before and after chemotherapy, as well as on the CT dataset. Future work will focus on model generalization across multi-centre datasets to enhance clinical applicability.