Comparison of Two-Dimensional- and Three-Dimensional-Based U-Net Architectures for Brain Tissue Classification in One-Dimensional Brain CT

Landscape of 2D Deep Learning Segmentation Networks Applied to CT Scan from Lung Cancer Patients: A Systematic Review

BACKGROUND

The increasing rates of lung cancer emphasize the need for early detection through computed tomography (CT) scans, enhanced by deep learning (DL) to improve diagnosis, treatment, and patient survival. This review examines current and prospective applications of 2D- DL networks in lung cancer CT segmentation, summarizing research, highlighting essential concepts and gaps; Methods: Following the Preferred Reporting Items for Systematic Reviews and Meta-Analysis guidelines, a systematic search of peer-reviewed studies from 01/2020 to 12/2024 on data-driven population segmentation using structured data was conducted across databases like Google Scholar, PubMed, Science Direct, IEEE (Institute of Electrical and Electronics Engineers) and ACM (Association for Computing Machinery) library. 124 studies met the inclusion criteria and were analyzed.

RESULTS

The LIDC-LIDR dataset was the most frequently used; The finding particularly relies on supervised learning with labeled data. The UNet model and its variants were the most frequently used models in medical image segmentation, achieving Dice Similarity Coefficients (DSC) of up to 0.9999. The reviewed studies primarily exhibit significant gaps in addressing class imbalances (67%), underuse of cross-validation (21%), and poor model stability evaluations (3%). Additionally, 88% failed to address the missing data, and generalizability concerns were only discussed in 34% of cases.

CONCLUSIONS

The review emphasizes the importance of Convolutional Neural Networks, particularly UNet, in lung CT analysis and advocates for a combined 2D/3D modeling approach. It also highlights the need for larger, diverse datasets and the exploration of semi-supervised and unsupervised learning to enhance automated lung cancer diagnosis and early detection.

Methods for improving colorectal cancer annotation efficiency for artificial intelligence-observer training

BACKGROUND

Missing occult cancer lesions accounts for the most diagnostic errors in retrospective radiology reviews as early cancer can be small or subtle, making the lesions difficult to detect. Second-observer is the most effective technique for reducing these events and can be economically implemented with the advent of artificial intelligence (AI).

AIM

To achieve appropriate AI model training, a large annotated dataset is necessary to train the AI models. Our goal in this research is to compare two methods for decreasing the annotation time to establish ground truth: Skip-slice annotation and AI-initiated annotation.

METHODS

We developed a 2D U-Net as an AI second observer for detecting colorectal cancer (CRC) and an ensemble of 5 differently initiated 2D U-Net for ensemble technique. Each model was trained with 51 cases of annotated CRC computed tomography of the abdomen and pelvis, tested with 7 cases, and validated with 20 cases from The Cancer Imaging Archive cases. The sensitivity, false positives per case, and estimated Dice coefficient were obtained for each method of training. We compared the two methods of annotations and the time reduction associated with the technique. The time differences were tested using Friedman's two-way analysis of variance.

RESULTS

Sparse annotation significantly reduces the time for annotation particularly skipping 2 slices at a time (P < 0.001). Reduction of up to 2/3 of the annotation does not reduce AI model sensitivity or false positives per case. Although initializing human annotation with AI reduces the annotation time, the reduction is minimal, even when using an ensemble AI to decrease false positives.

CONCLUSION

Our data support the sparse annotation technique as an efficient technique for reducing the time needed to establish the ground truth.

Accuracy of skull stripping in a single-contrast convolutional neural network model using eight-contrast magnetic resonance images

In automated analyses of brain morphometry, skull stripping or brain extraction is a critical first step because it provides accurate spatial registration and signal-intensity normalization. Therefore, it is imperative to develop an ideal skull-stripping method in the field of brain image analysis. Previous reports have shown that convolutional neural network (CNN) method is better at skull stripping than non-CNN methods. We aimed to evaluate the accuracy of skull stripping in a single-contrast CNN model using eight-contrast magnetic resonance (MR) images. A total of 12 healthy participants and 12 patients with a clinical diagnosis of unilateral Sturge-Weber syndrome were included in our study. A 3-T MR imaging system and QRAPMASTER were used for data acquisition. We obtained eight-contrast images produced by post-processing T1, T2, and proton density (PD) maps. To evaluate the accuracy of skull stripping in our CNN method, gold-standard intracranial volume (ICVG) masks were used to train the CNN model. The ICVG masks were defined by experts using manual tracing. The accuracy of the intracranial volume obtained from the single-contrast CNN model (ICVE) was evaluated using the Dice similarity coefficient [= 2(ICVE ⋂ ICVG)/(ICVE + ICVG)]. Our study showed significantly higher accuracy in the PD-weighted image (WI), phase-sensitive inversion recovery (PSIR), and PD-short tau inversion recovery (STIR) compared to the other three contrast images (T1-WI, T2-fluid-attenuated inversion recovery [FLAIR], and T1-FLAIR). In conclusion, PD-WI, PSIR, and PD-STIR should be used instead of T1-WI for skull stripping in the CNN models.

Deep Learning Based Detection of Enlarged Perivascular Spaces on Brain MRI

Deep learning has been demonstrated effective in many neuroimaging applications. However, in many scenarios, the number of imaging sequences capturing information related to small vessel disease lesions is insufficient to support data-driven techniques. Additionally, cohort-based studies may not always have the optimal or essential imaging sequences for accurate lesion detection. Therefore, it is necessary to determine which imaging sequences are crucial for precise detection. This study introduces a deep learning framework to detect enlarged perivascular spaces (ePVS) and aims to find the optimal combination of MRI sequences for deep learning-based quantification. We implemented an effective lightweight U-Net adapted for ePVS detection and comprehensively investigated different combinations of information from SWI, FLAIR, T1-weighted (T1w), and T2-weighted (T2w) MRI sequences. The experimental results showed that T2w MRI is the most important for accurate ePVS detection, and the incorporation of SWI, FLAIR and T1w MRI in the deep neural network had minor improvements in accuracy and resulted in the highest sensitivity and precision (sensitivity =0.82, precision =0.83). The proposed method achieved comparable accuracy at a minimal time cost compared to manual reading. The proposed automated pipeline enables robust and time-efficient readings of ePVS from MR scans and demonstrates the importance of T2w MRI for ePVS detection and the potential benefits of using multimodal images. Furthermore, the model provides whole-brain maps of ePVS, enabling a better understanding of their clinical correlates compared to the clinical rating methods within only a couple of brain regions.

Automatic segmentation of the core of the acoustic radiation in humans

Introduction

Acoustic radiation is one of the most important white matter fiber bundles of the human auditory system. However, segmenting the acoustic radiation is challenging due to its small size and proximity to several larger fiber bundles. TractSeg is a method that uses a neural network to segment some of the major fiber bundles in the brain. This study aims to train TractSeg to segment the core of acoustic radiation.

Methods

We propose a methodology to automatically extract the acoustic radiation from human connectome data, which is both of high quality and high resolution. The segmentation masks generated by TractSeg of nearby fiber bundles are used to steer the generation of valid streamlines through tractography. Only streamlines connecting the Heschl's gyrus and the medial geniculate nucleus were considered. These streamlines are then used to create masks of the core of the acoustic radiation that is used to train the neural network of TractSeg. The trained network is used to automatically segment the acoustic radiation from unseen images.

Results

The trained neural network successfully extracted anatomically plausible masks of the core of the acoustic radiation in human connectome data. We also applied the method to a dataset of 17 patients with unilateral congenital ear canal atresia and 17 age- and gender-paired controls acquired in a clinical setting. The method was able to extract 53/68 acoustic radiation in the dataset acquired with clinical settings. In 14/68 cases, the method generated fragments of the acoustic radiation and completely failed in a single case. The performance of the method on patients and controls was similar.

Discussion

In most cases, it is possible to segment the core of the acoustic radiations even in images acquired with clinical settings in a few seconds using a pre-trained neural network.