The most prevalent cerebral tumors are brain metastases (BMs), which frequently develop from aggressive melanoma, breast cancer, and lung cancer [
1]. Due to more comprehensive therapeutic options and lung cancer monitoring initiatives in many nations, cancer patients are living longer, which has increased the incidence of this disease. The diagnosis of BMs relies heavily on the use of contrast-enhanced T1-weighted imaging (CET1WI) magnetic resonance (MR) scans, which are also employed for long-term monitoring to gauge therapy effectiveness [
2]. The majority of patients have three or fewer brain tumors when they first appear, although 40% of patients have more than three. Despite the fact that identifying BMs is a laborious and time-consuming physical process for radiologists, they are very important in the preliminary diagnosis of tumors, definition of the initial tumor volume, and monitoring of volume fluctuations as a result of therapy [
3]. The early and precise identification of BMs is essential for proper therapy planning, as the existence of BMs can shift overall oncologic management. To aid physicians, deep learning-based techniques have recently been proposed for autonomous spotting or segmenting of BMs in MRI images [
4]. The fact that BM and other structures, such as cerebral arteries, share comparable physical characteristics, and that BMs vary greatly in size and spread, renders this a difficult task [
5]. Recently, many researchers have introduced multi-sequence MRI automated recognition and segmentation methods circumvent the constraints of using only MRI sequences. According to this perspective, precise BM identification and differentiation from various suspect areas (BM imitators) are crucial for effective evaluation and treatment [
6]. Early and precise detection of the cause(s) of BM prior to surgery could alter individual treatment strategies, which is of great therapeutic significance. The epidermal growth factor receptor (EGFR) [
7] gene mutant state is critical for treatment plans, such as EGFR-tyrosine kinase inhibitor (EGFR-TKIs) administration, for NSCLC patients with BMs. The HER2 condition is essential to deciding on treatment approaches for BC patients with BMs. This is due to the fact that patients who are HER2-positive frequently receive tailored antibody treatments and typically have poor prognoses [
8]. Since it may not always be practically feasible to obtain tissue from the main tumor, the spread may serve as a valuable backup source of information regarding the features and gene status of the primary tumor. However, due to the lack of particular indicators, radiologists are hardly ever able to determine the spread sources or examine MRI images visually to determine the gene state of the primary tumor. Artificial intelligence (AI) and computer-aided diagnostic (CAD) developments are becoming ever more significant in the area of medical imaging [
9]. The term ‘radiomics’ refers to the methodical calculation and study of a significant number of numeric characteristics used in medical imaging. However, few investigations have used CAD to identify the spreading cause of BM. In earlier studies, it was assumed that the total brain spread of tumor volume was uniform. However, recent studies have shown that solid tumors can be diverse, with several tumor areas being more physiologically invasive and possibly reflecting various biological processes [
10]. Intra-tumor heterogeneity, or ITH, has been recognized for having important ramifications that represent unique tumor development. To segment the entire tumor area into intra-tumor sub-regions and enable the collection of useful information from the sub-regions, sub-region-based radiomics methods have been proposed. According to statistics, sub-region radiomics studies have been carried out in cases of esophagus carcinoma, lung cancer, and breast cancer, and have demonstrated the potential to greatly enhance the diagnostic performance of radiomics techniques [
11]. To the authors’ understanding, brain cancer has not been studied using sub-region radiomics. Therefore, the following are the accomplishments of this work:
This simplified form of the U-Net can be readily implemented on devices with limited (processing and memory) resources. This research is novel in several ways, as it employs the shallowest version of U-Net on a large dataset to derive characteristics that improve the BP forecast process.
The method known as HybWWoA, which combines the water waves optimization and whale optimization techniques, is used to choose the features by reducing the size of the recovered feature.
DenseNet was developed to use dense connections from its backbone design to identify BM MRI images with high precision and accelerated learning.