Lung cancer is the most frequent source of brain metastases (BMs), and 30%–50% of patients with lung cancer develop BMs during the course of the disease (1). As a result, brain magnetic resonance imaging (MRI) has become an important part of staging and treatment planning for lung cancer. Many guidelines recommend brain MRI for the screening and follow-up of BMs in advanced non-small lung cancer or small cell lung cancer (2, 3). However, the detection of small BMs and an accurate assessment of treatment response require tedious effort by radiologists. In addition, stereotactic radiosurgery has become popular in the treatment of BMs; therefore, manual segmentation of BMs has significantly increased the workload of radiosurgeons (4, 5).

In this context, recent studies (6–8) have implemented deep learning models, particularly deep convolutional neural networks (CNNs), for the automatic detection and segmentation of BMs, and have reported promising results with sensitivities of up to 90% and Dice coefficients of up to 0.8. However, the studies often report a substantial number of false-positive (FP) results and low sensitivity in detecting small BMs. Moreover, their segmentation methods were based on multiparametric scans such as the T1-weighted image T2-weighted image (T2WI), contrast-enhanced T1WI, and fluid-attenuated inversion recovery (FLAIR). However, these methods are not always favorable because additional sequences may increase the scan time and are often acquired with a larger thickness and lower resolution, which may add uncertainty to the segmentation. A few studies (9, 10) have used a single modality—in particular, the contrast-enhanced 3D gradient echo (GRE) T1WI sequence. However, recent studies (11, 12) have demonstrated that the three-dimensional (3D) black blood (BB) T1WI sequence is superior to the 3D GRE T1WI sequence in detecting small BMs by suppressing intraluminal blood signals. In a subsequent study, deep learning (DL)-based methods for BM detection and segmentation, utilizing the 3D BB T1WI sequence, demonstrated a better performance advantage over methods employing the 3D GRE T1WI sequence (8). In that study, the sensitivity for detecting brain metastases (BM) on 3D BB T1WI was higher at 92.6% compared to the sensitivity on 3D GRE T1WI, which stood at 76.8%.

Another limitation of previous studies is that internal necrosis was included in the BM segmentation. BM necrosis may represent a by-product of chemotherapy or radiation therapy (13, 14). The Response Assessment in Neuro-Oncology Brain Metastases (RANO-BM) criteria also recommend that these necrotic or cystic cavities should not be measured for determining a response (15). Thus, previous BM segmentation algorithms that included solid components and necrosis may lead to inappropriate treatment assessment.

The aims of our study were two-fold (1): to assess whether a DL algorithm using a single modality, 3D BB T1WI, has promising performance for the detection and segmentation of BMs and (2) to investigate whether the volumetric assessment using our developed DL algorithm, excluding necrosis, is comparable to the conventional assessment based on the RANO-BM criteria.

RLK-Unet for the detection and segmentation of BMs has two clinically favored features that previous models have rarely tried. First, RLK-Unet was based only on a single modality, 3D BB T1WI. Second, RLK-Unet segments the solid part of the tumor to avoid necrosis. Nevertheless, RLK-Unet exhibited promising performance for detection and segmentation. Moreover, the volumetric assessment by RLK-Unet strongly agreed with that of the response assessment by the radiologist, based on the RANO-BM criteria. Thus, our model is expected to facilitate clinical workflow and to potentially improve patient outcomes via a volumetric assessment of the treatment response.

While earlier studies demonstrated high sensitivity in the detection of BMs, surpassing 80%, they were accompanied by a significant number of FPs and, consequently, exhibited low precision, as indicated in Table 5 (7, 10, 28, 29). Subsequent research, incorporating multiple modalities, showed improved performance with sensitivities ranging from 82% to 100% and reducing the FP rate to between 0.6 and 1.5 per scan (6, 8, 30). Notably, a recent study, utilizing a single modality, introduced a novel loss function and integrated temporal prior information, achieving exceptional results (sensitivity: 84%; precision: 99%; FP rate: 1) (31). Another extensive study also reported remarkable outcomes (sensitivity: 88.4%; precision: 90.1%; FP rate: 0.4) (32). However, it is essential to acknowledge that most of these studies did not address the critical issue of excluding internal necrosis, which is pivotal for accurate volumetric assessment of tumor burden. In contrast, our RLK-Unet successfully addressed the exclusion of necrotic regions within BMs, achieving outstanding performance (sensitivity: 86.9%; precision: 79.6%; FP rate: 1.8). To achieve this, we implemented several strategies within our DL algorithm, enabling us to maintain high sensitivity while concurrently reducing the FP rate.

First, the DL model was based on a BB image. A previous meta-analysis (33) reported the superiority of BB images for the detection of small BMs (<5 mm) because these images suppress the blood signal and have a higher contrast-to-noise ratio, compared to GRE images. In accordance with this finding, RLK-Unet maintained a high sensitivity of 80.84 in detecting small BMs (≤10 mm), whereas previous models showed a relatively lower performance for small BMs (sensitivity: 15–50) (7, 29). Second, we used a few large kernels instead of a stack of small kernels in the CNN. This approach resulted in larger effective receptive field more efficiently, thereby significantly increasing the sensitivity from 84.52 to 88.36 ( Supplementary Material S2 ) (23). However, because of trade-off between sensitivity and precision, the precision of RLK-Unet was unfortunately decreased from 80.6 to 68.4. To replenish this, we implemented MHFs, which maximize the contrast between BMs and normal brain tissue, thereby increasing precision. Lastly, the surface mask effectively decreased FPs, by suppressing some blood vessels that were incompletely suppressed in BB images (34). The choroid plexus also frequently mimicked BMs in our model. It was successfully removed using the choroid plexus mask.

RLK-Unet demonstrated a DSC of 0.66 in segmenting BMs. This value is lower than that reported in previous studies (0.77–0.85) (7, 10, 28). We suggest the following explanation for this result: the DSC cannot incorporate the size of the BMs within its score. Only small pixel differences between the GT and the prediction in small BMs may substantially decrease the score ( Figure 3 ) (35). In line with this suggestion, our results showed excellent segmentation performance in larger BMs (DSC of large BMs vs. small BMs: 0.85 vs. 0.54). We presume that small pixel differences in the segmentation of small BMs rarely affect the volumetric assessment. The excellent agreement in the volume measurement of the BM between the GT and the prediction in our results also supports our assumption.

Volumetric measurement may provide a more objective and sensitive quantification to evaluate tumor response to treatment than does linear measurement in the current RANO-BM criteria (36). However, it is not clinically feasible because the manual volumetric measurement is a labor-intensive, time-consuming, and complex task (37). The clinical significance of our work lies in the fact that our automated DL algorithm may alleviate these tedious and labor-intensive tasks while maintaining results similar to those of conventional tumor assessment by a radiologist. Cho et al. (38) recently showed the possibility of end-to-end automated treatment response evaluation of BM. However, the sensitivity of BM detection in their system was relatively low (58.0%–80.0%). In addition, their BM segmentation method included internal necrosis, which should be avoided in volumetric measurements. Previous studies have reported that the presence of necrosis in BMs may be an indication of a response to chemotherapy or radiation therapy (14). Furthermore, various imaging characteristics can change during the course of treatment. For instance, patients receiving a combination of tyrosine kinase inhibitors and intracranial radiation therapy are more likely to experience hemorrhages within their BMs (39). Additionally, the values of the apparent diffusion coefficient show alterations before and after chemoradiation therapy (40). As a result, monitoring changes in these imaging characteristics is essential for assessing the treatment effects on BMs. Considering these aspects, our method may offer improved performance and better alignment with real-world clinical scenarios. Based on these perspectives, our method may have better performance and may better reflect real-world clinical settings.

However, RLK-Unet also showed three disagreements with the conventional RANO-BM criteria for treatment assessment (5.1%; 3/58 patients). RLK-Unet may overestimate treatment responses because it records an equivocal enhancement as a true lesion and may underestimate treatment responses because it ignores subtle enhancement after treatment. The incorporation of dynamic information from longitudinal images into the DL algorithm may improve performance. With an in-depth comparison of pre- and posttreatment images, the DL algorithm may better detect subtle changes in tumor size and assess the treatment response more precisely (41).

Our study has some limitations. First, it was a retrospective single-center study, which is insufficient to address variability in scanning techniques and hardware implementation across hospitals. We used five-fold cross-validation for detection and segmentation and a temporally separated internal test set for the treatment response assessment; however, a multicenter study in the near future is required to improve the generalizability of our results. Second, RLK-Unet has some limitations in assessing leptomeningeal seeding, pachymeningeal seeding, and skull metastases because we excluded these factors from our cohort or removed the skull during preprocessing. Third, RLK-Unet was based on patients with lung cancer and may not be applicable to patients with other primary cancers. Finally, in this work, a contrast-enhanced BB T1WI (3D fast spin echo T1-weighted technique) was used for developing our algorithm because a previous study showed that the performance of an algorithm based on 3D BB T1WI was superior to that based on 3D GRE T1WI (sensitivity: 92.6 vs. 76.8) (8). Our study aligns with this result, with sensitivity, DSC, and precision for 3D BB T1WI and 3D GRE T1WI as follows: 86.9, 0.66, 79.6 vs. 53.7, 0.46, 68.7, as shown in Supplementary Material 7 . Consequently, our algorithm may not be optimally applied to the 3D GRE T1WI sequence, which is more widely used for BM imaging. Lastly, the performance of our algorithm may not be directly compared with previous studies because of a different dataset. However, we ran publicly available algorithms such as 3D U-Net and nnU-Net, which were utilized in prior studies (8, 30, 32), for our dataset, and their performances are inferior to the results of our algorithm ( Supplementary Material S7 ). Consequently, we may conclude that RLK-Unet shows a comparative performance for BM detection and segmentation.

Our developed DL model for the treatment response assessment of BM had more favorable features in clinical practice than did models reported in previous studies. RLK-Unet uses a single modality but shows excellent performance for the detection and segmentation of BMs, even for small metastases. Moreover, our segmentation results very well predicted GT, while avoiding cysts or necrosis, and exactly measured the volumetric tumor burden. The assessment of the treatment response showed good agreement with the decision of the radiologists. We believe that this research takes DL-based BM evaluation to the next level and may facilitate the clinical workflow for radiologists or neuro-oncologists.

The original contributions presented in the study are included in the article/ Supplementary Material . Further inquiries can be directed to the corresponding authors.

The studies involving humans were approved by Gangnam severance hospital IRB. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin because it was a retrospective study. Written informed consent was not obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article because our institutional review board waived the requirement for informed consent because it was a retrospective study.

SS: Formal Analysis, Writing – original draft. BJ: Data curation, Writing – review & editing. MP: Writing – review & editing. SS: Supervision, Writing – review & editing. HO: Data curation, Methodology, Writing – review & editing. JK: Methodology, Supervision, Writing – review & editing. SL: Project administration, Writing – review & editing. SA: Conceptualization, Funding acquisition, Supervision, Writing – original draft, Writing – review & editing. J-ML: Conceptualization, Funding acquisition, Writing – review & editing.