Deep learning-based detection and quantification of brain metastases on black-blood imaging can provide treatment suggestions: a clinical cohort study

Objectives We aimed to evaluate whether deep learning–based detection and quantification of brain metastasis (BM) may suggest treatment options for patients with BMs. Methods The deep learning system (DLS) for detection and quantification of BM was developed in 193 patients and applied to 112 patients that were newly detected on black-blood contrast-enhanced T1-weighted imaging. Patients were assigned to one of 3 treatment suggestion groups according to the European Association of Neuro-Oncology (EANO)-European Society for Medical Oncology (ESMO) recommendations using number and volume of the BMs detected by the DLS: short-term imaging follow-up without treatment (group A), surgery or stereotactic radiosurgery (limited BM, group B), or whole-brain radiotherapy or systemic chemotherapy (extensive BM, group C). The concordance between the DLS-based groups and clinical decisions was analyzed with or without consideration of targeted agents. The performance of distinguishing high-risk (B + C) was calculated. Results Among 112 patients (mean age 64.3 years, 63 men), group C had the largest number and volume of BM, followed by group B (4.4 and 851.6 mm3) and A (1.5 and 15.5 mm3). The DLS-based groups were concordant with the actual clinical decisions, with an accuracy of 76.8% (86 of 112). Modified accuracy considering targeted agents was 81.3% (91 of 112). The DLS showed 95% (82/86) sensitivity and 81% (21/26) specificity for distinguishing the high risk. Conclusion DLS-based detection and quantification of BM have the potential to be helpful in the determination of treatment options for both low- and high-risk groups of limited and extensive BMs. Clinical relevance statement For patients with newly diagnosed brain metastasis, deep learning–based detection and quantification may be used in clinical settings where prompt and accurate treatment decisions are required, which can lead to better patient outcomes. Key Points • Deep learning–based brain metastasis detection and quantification showed excellent agreement with ground-truth classifications. • By setting an algorithm to suggest treatment based on the number and volume of brain metastases detected by the deep learning system, the concordance was 81.3%. • When dividing patients into low- and high-risk groups, the sensitivity for detecting the latter was 95%. Supplementary Information The online version contains supplementary material available at 10.1007/s00330-023-10120-5.


MRI acquisition protocols
A black-blood imaging sequence was added to our routine BM protocol in accord with the consensus recommendations for MRI of brain metastasis [1].Five minutes after administration of gadolinium-based contrast (0.05 ml/kg gadoterate meglumine; Dotarem; Guerbet), gradient-echo(GRE) CE-T1WI and black-blood CE-T1WI were acquired.excitation pulse with motion-sensitized gradients between radiofrequency pulses.The duration between the two 90° pulses was 28.3 ms, and the flow velocity encoding for gradient pulses was 3 cm/s.

DLS for detection and quantification of BM
The GRE CE-T1WI and black-blood CE-T1WI were coregistered using rigid transformations with six degrees of freedom in SPM (version 12, www.fil.ion.ucl.ac.uk/spm/).
Skull stripping was performed using an algorithm optimized for heterogeneous MRI data with

Discrepancies between actual clinical decisions and DLS-based suggestions
Group A was indicated for follow-up, group B was limited to patients with BMs indicated for surgery or SRS, and group C was patients with extensive BMs suggested for chemotherapy or WBRT by the DLS.The case showing the greatest discrepancy was a patient who was suggested as belonging to group A by the DLS, whereas the patient actually received WBRT.The patient had pancreatic cancer and two BMs in the cerebellum with a total volume of 27 mm 3 .He also had disseminated bone metastasis in the whole spine.The clinician initially planned SRS for BMs; however, the delay for the scheduled SRS was considered too long, and radiation therapy was performed simultaneously on the spine and cerebellum.
Four patients were suggested as group B by the DLS but did not receive any treatment.
Of these, one patient had primary lung cancer and five BMs with a total volume of 155 mm 3 .
The burdens from primary lung cancer and liver metastasis were high at the same time.
Although she needed treatment for BM, the treatments for lung and liver lesions were given clinical priority, and the patient was initially followed-up, with SRS for BM then being performed 7 months later.The other three patients had NSCLC with PDL1 expression and underwent pembrolizumab treatment as part of a clinical trial.
The last two cases (patients 20 and 21) in the table were considered likely to benefit from DLS-based treatment.Although they had multiple lesions, only selective stereotactic radiosurgery was performed on a few lesions, and follow-up MRI showed an increase in the size and number of the other lesions.We believe that if whole brain RT had been performed, as suggested by the DLS, different results may have been obtained.

Supplementary
diverse pathology or post-treatment changes (https://github.com/MIC-DKFZ/HD-BET).The lesion segmentation model was implemented using nnUnet, a 3D U-Net-based method (https://github.com/MIC-DKFZ/nnUNet)[2] in the PyTorch package version 1.1 in Python 3.7 (www.python.org).In our model training, a GRE CE-T1WI and black-blood CE-T1WI image pair were fed into the model as input.A full-resolution 3D model was developed because it was expected that such a model would show optimum performance with the small simpleshaped lesions characteristic of BM and the possibility of multiple lesions within a single image.The training data were augmented with rotation, gamma, scaling, elastic deformation, and mirror transformations.The model training details included dice and cross-entropy loss function, Adam optimizer with learning rate weight decay, learning rate (initial) 0.0003, batch size 2, and patch size 128 × 128 × 112.Model training was conducted with an NVIDIA TITAN RTX 24 GB GPU with CUDA version 10.0, and the model performance converged after 507 epochs over 2.5 days.

Table 1 .
Clinical characteristics and distributions of brain metastases in the training set