Doctor Zhang Nan
《Medicine (Baltimore)》 was published online on October 7, 2022《Medicine (Baltimore)》》》《Using artificial neural network to predict brain metastases after radiosurgery treatment by Hyeong Cheol Moon and Young Seok Park of Chungbuk National University Hospital in South Korea. Volume prediction for large brain metastases after hypofractionated gamma knife radiosurgery through artistic neural network》 (doi: 10.1097/MD.0000000000000030964.). The effectiveness of single-time gamma knife radiosurgery (GKRS) in the treatment of cerebellar metastasis has been confirmed, but the effectiveness of large-segment GKRS (hfGKRS) based on the linear quadratic equation (LQ) model for the treatment of brain metastasis (BM) is uncertain. The purpose of this study is to study the volume changes of large BM after hfGKRS from the LQ model, and to predict volume changes using the artificial neural network (ANN).
Brain metastases (BM) are usually considered to be intracranial tumors with concurrent systemic cancers and are a key cause of morbidity and death in patients. The increase in the incidence of BM may be due to better control of local tumors through surgery, radiotherapy, radiosurgery and systemic chemotherapy, thereby prolonging survival. Recently, stereotactic radiosurgery (SRS), such as gamma knife radiosurgery (GKRS), has become increasingly popular as the treatment of metastatic brain tumors. However, it has been reported in some studies that SRS treatment of large lesions (3 cm in diameter) is still difficult to control due to radiation toxicity relative to single-course SRS. Previous reports have rarely confirmed the use of GKRS for large BM treatment.
The new generation of ICONTM gamma knife with mask fixation has the potential for segmentation therapy. ICON gamma knife can detect the patient's motion trajectory through high-definition motion management (HDMM) camera, define stereotactic references using conical beam computed tomography (CBCT), and may use mask fixation to achieve large segmentation therapy. Although large segmentation GKRS (hfGKRS) treatment based on linear quadratic (LQ) models can effectively treat large brain metastases, optimization parameters including treatment cycles, dose prescriptions, and volumetric results are not known. In recent years, statistical and mathematical models of clinical decision-making have become an important area of concern to researchers. Models in the clinical field can help doctors make decisions, optimize treatment options, and prevent the development of risk factors.
Artificial neural network (ANN) and regression models are used to predict results. Artificial neural networks are a powerful analyzer that discovers complex and nonlinear relationships between data sets and mimics biological nervous systems. The introduction of artificial neural networks is one of the main challenges in developing the prediction model in the field of radiosurgery. Gradient-based algorithms are the most frequently trained algorithms.
In this study, we studied the volume changes of large BM after receiving hfGKRS treatment from the LQ model. Furthermore, we used artificial neural networks to predict volume changes.
After obtaining the Institute’s ethical permission, we retrospectively created data at our ICON Gamma Knife Center to analyze large BM patients from 2018 to 2021. A total of 28 patients (18 males and 10 females; age range 35-85 years; median age was 69.5 years), and were diagnosed with large BM at Chungbuk National University Hospital. General characteristics such as gender, age, pathology, recursive partition analysis (RPA), Karnofsky general performance status (KPS) score, and GKRS operation report. We analyzed 44 large brain lesions in 28 patients.
All subjects were subjected to magnetic resonance imaging (MRI) scan using a 1.5T MR system (Philips Achieva, Best, Netherlands). Acquisition of continuous 3D T1 weighted enhancement images (layer thickness 1.0 mm; repeat time/echo time, 25/46.2 ms; flip angle, 30; field of view 256x256 and 240x240 matrix; section number, 80; acquisition time, 360 - 420 seconds). We also added T1-weighted and T2-weighted images for diagnosis. Exclude all images with motion artifacts.
is based on T1 weighted enhanced images and determines the number of large brain metastases at the time of referral with the consent of medical physicist/neurosurgeon/radiologist trained in medical imaging and neurooncology. All patients need MRI follow-up after at least about 3 months.Tumor volume was measured by Leksell Gamma Plan (Elekta Instrument AB, Version 11.1) using manual and semi-automatic segmentation.
feature extraction was performed from electronic medical records and planning parameters: gender, age, KPS, RPA, number of targets, beam-on time, coverage, selectivity, gradient index, prescription dose (50% edge dose), number of segmentation, bioequivalent dose (bio-effect dose [BED]10), number of lesions, diagnosis, initial tumor volume, lesion area and tumor volume after hfGKRS treatment. Study on the effect of segmentation treatment of BED10 using LQ model. The lesion area is defined by the main cerebral lobe (frontal lobe, temporal lobe, parietal lobe, occipital lobe and cerebellum).
We used Kerash (version 2.2.4), Pandash (version 0.23), and scikit-learn python (version 3.3) libraries. The single-layer perceptron (SLP) is composed of a single-layer output node, and the output is directly caused by the weight and deviation. The multi-layer perceptron (MLP) consists of a feed-forward algorithm (called function approximation in many applications) and a backpropagation network. Our training process is optimized using the Keras tool for classification cross entropy loss and stochastic gradient descent. Learning rate is 1.0. The ANN modeling flowchart is shown in Figure 1. The characteristics of
28 patients who received hfGKRS due to large BM are listed in Table 1. Follow-up every 3 months to observe recurrence or new tumors. The initial tumor volume was 14.0±5.3 cc, and the tumor volume was 9.0±4.3 cc after hfGKRS treatment. Except for 2 exceptions, all the large tumors had radiographic reactions.
When planning GKRS, the tumor edge was 0.5 mm larger than the existing tumor resection edge, as shown in Figure 2. We simply calculated prescription doses (50% edge) from the LQ model using MATLAB (R2018a version, MathWorks, USA). Almost all BMs follow α/β = 10. According to the LQ model equation (BED = nD [1 + [D/ (α/β]]), the prescription dose was selected based on the number of segmentations in Figure 3.
For calculation accuracy, we converted the brain tumor size into an integer to obtain the best selection performance. In the test set, the accuracy of MLP and SLP in predicting volume changes was 80% and 75%, respectively.
followed up for 3 months to 12 months. The local tumor control rate was 96%, and no new brain metastasis was found in all patients. The median overall survival was 6 months. The clinical courses of neurological deficit include headache (n = 3), motor dysfunction (n = 1), and vomiting (n = 1). Follow-up MRI data showed that all patients had improvement.
We retrospectively investigated the clinical manifestations of 28 patients with large (3 cm diameter or 10 cc in volume) brain metastases (BM) treated with hfGKRS. A total of 44 tumors were extracted from 28 patients with above characteristics. 30 large brain metastases were randomly assigned as training sets and 14 large brain metastases were used as test sets. To predict volume changes after hfGKRS, we used artificial neural network models (ANN models) (single-layer perceptron, SLP) and multi-layer perceptron [multi-layer] perceptron, MLP]).
According to the LQ model, large BMs reduced volume by 96% after receiving hfGKRS treatment. The prediction accuracy of SLP and MLP in the ANN model is 70% and 80% respectively. Even in large brain metastases (BMs), hfGKRS in the LQ model may be a good treatment option. In addition, the MLP model can predict volume changes after receiving hfGKRS treatment with an accuracy of 80%.
This study aims to study volume changes after treating large BMs in hfGKRS and use artificial neural networks to predict volume changes. The main results of our case study are that the LQ model can be applied to hfGKRS treatment, The hfGKRS series can reduce tumor size by large BM, and ANN can predict tumor volume changes with an accuracy of 80%.
Artificial Neural Network (ANN) is the most popular artificial intelligence technology in medicine. ANN has been used in clinical diagnosis, image analysis, data interpretation, neuro-oncology and histopathology in clinical diagnosis, radioactive effects, ANN has similar capabilities to computers, can collect and process many variables, and has the ability to train trial and error. Therefore, computers can learn to recognize patterns and make informed decisions. This technology is called Artificial Intelligence , using variable technology in the medical field.However, ANN does not support unique solutions because the resting state of training is based on several factors, including weights, case counts, and test cycles. Therefore, for some applications, such as cancer prediction, a frequency distribution of the network relative to the probability of the result can be generated and a central trend can be created including mean, pattern, variance measurements and non-parametric prediction intervals (for non-parametric distributions with slopes). The LQ model was used to describe the cell survival curve, which consists of two mechanisms of radiation-induced cell death. The purpose of hfGKRS is to provide optimal doses for large-volume metastases rather than traditional radiotherapy while minimizing damage to normal tissue. Iwata et al. argue that LQ formalism leads to a wrong large-segment radiotherapy model, because the efficacy of large-segment is about 15%. Furthermore, the α/β ratio of metastatic brain tumors was assumed to be 10-20, and the higher α/β ratio indicated more sensitive to segmentation therapy. Using the LQ model, the clinical results of hfGKRS in the treatment of metastatic brain tumors have not been optimized. We found that MLP can predict about 80% of clinical outcomes. These authors believe that neural network can be used as an optimized treatment planning method for predicting clinical outcomes. We adopt a daily segmentation treatment plan. Some daily segmentation treatments have been reported. Shoji et al. irradiated 20 to 30 Gy in two segments from 3 to 4 weeks. Kim et al. used a head rack for 5-11Gy segmentation treatment for 3-4 days. Dohm et al. irradiated 14Gy/1fx one month after 15Gy/1fx. The purpose of these interval-time strategies is to reduce the tumor size after a second re-planning of smaller volumes. In our current study, the volume was reduced to 96%, and no radionecrosis was seen. We believe that daily treatment plans for large BMs are effective.
This study has certain limitations. First, the number of patients with brain tumors is relatively small, and future studies require more samples. Second, the study shows the cross-section of brain tumor development. Third, there is no comparison of radiation responses to the primary cancer type.
We analyzed the effect of hfGKRS on large BM and predicted the change in volume through neural networks. The results show that the hfGKRS algorithm based on the LQ model is suitable for large BMs. Through our neural network model, the volume changes of large BMs after hfGKRS treatment can be predicted.
