Lung cancer is one of the major challenges in medical oncology. It is the most common cause of cancer related death in the world [1–3]. Up to 64% of all patients with lung cancer develop brain metastasis [4, 5]. Data from several patient series have shown that the median age of anaplastic lymphoma kinase (ALK) positive non-small cell lung cancer (NSCLC) patients is 55 years which is about 10 years to 15 years lower than for the general NSCLC population and also the epidermal growth factor receptor (EGFR) mutated patients, approximately 70% of these patients are never smokers [6]. Human genome sequencing has ensured identification of epigenetic mutations, tumour-suppressor-gene inactivation and pro-oncogene mutations that might be potentially targeted for therapy [7–12]. For NSCLC, EGFR mutations and on abnormal fusion of ALK being targeted successfully with EGFR tyrosine kinase inhibitors (TKI) and ALK inhibitors respectively.

Classically, brain metastases seen on magnetic resonance imaging (MRI) occurring at grey and white matter junction can have varying imaging parameters such as size, post contrast enhancement, peri-lesional edema. There has been increased interest in research for the characterization of quantitative imaging features reflecting tumor biology, physiology and tumor phenotype using artificial intelligence (AI) based algorithms. Radiomics and deep learning (DL)-AI based models are extensively used in medical imaging [13–15] which basically involves image segmentation and detection. This image segmentation usually relies on manual delineation which is time-consuming and subject to inter- or intra-segmentation variation. Secondly, despite the segmentation of images being accurate, no standard method for image feature extraction is available and errors due to miscalculation are encountered as it is difficult to verify the accuracy and reproducibility of image features.

Advanced AI models such as the neural network based DL method can overcome these problems through a self-learning strategy and presents a promising tool for genomic analysis [16–19]. The DL method can be correlated with the functioning of the neural network in the brain. In comparison to radiomic methods, precise tumor boundary annotation is not required with DL and thus saves a lot of time and human effort [20–24]. Furthermore, DL method takes into consideration the microenvironment of the surrounding lung parenchyma and can extract features that are adaptive to specific clinical outcomes, whereas radiomics can only describe general features which lacks specificity for outcome prediction [25–29].

The primary objective of this study was to develop a novel DL model to segment brain metastasis and to categorize the lesions into EGFR positive, ALK positive and double negative groups. Additionally, classification was also performed based on semantic features i.e., clinical and imaging features.

A significant proportion of patients with primary lung cancer develop brain metastasis during the disease course. Recent developments in the research of pathophysiology and molecular biology have led to the further understanding of different molecular mutations, which have been proven potentially useful in therapeutic oncology [30–32]. EGFR mutation and ALK rearrangement have been in focus recently since they can be targeted for molecular therapy. Recently developed and more specific molecular targeted therapy such as osimertinib in EGFR mutation [33] has been proven to increase the survival in patients with a specific mutation.

In the first part of the study, 117 patients were analysed of carcinoma lung with positive MRI scan for brain metastasis in which clinical parameters showed statistical significance. In this study, clinical parameters viz., age and smoking history showed a statistically significant impact on mutation status (P = 0.001 and P = 0.0002, respectively). A clinical study conducted by Gao et al. [34] showed age and smoking history have a statistically significant impact on ALK status. Another clinical study by Bhatt et al. [35] showed age and sex had a statistically significant impact on EFGR mutation (P = 0.001 and P = 0.0001, respectively). Similarly, a study by Rangachari et al. [36] statistically significant impact of age on EFGR mutation (P = 0.0001). However, it should be noted that above-mentioned studies did not use imaging parameters. Statistical analysis of semantic features in the current study showed both EGFR mutation and ALK rearrangement were significantly less prevalent in smokers who developed NSCLC. Fuzzy T2 borders and intralesional haemorrhage could be used to differentiate EGFR positivity from the double negative group. Ring enhancement, meningeal involvement and synchronous lung metastasis can be used to differentiate ALK positivity from the double negative group. Peripheral diffusion restriction, ring enhancement and T2W hypo intense signal intensity can be used to differentiate ALK positivity from EFGR positivity.

In the second part of the study, performance of the DL model for classification of brain metastases was evaluated. A deeper understanding of radiomics has shown much finer imaging features specific to a mutation can be delineated even when using standard imaging sequences. But consistent manual identification of such features might not always be accurate and is more time consuming. Development of an automated computed algorithm, which can be trained to identify the subtle subclinical imaging features with precision and within a much shorter timeframe, was the purpose of this study. Pre-treatment basic MR sequences namely T1, T2, FLAIR and T1post contrast sequences of 117 patients having brain metastasis belonging to EGFR positive, ALK positive and double negative groups were included. The case parameters were derived from the studies by Wadhwa et al. [37] and Mahajan et al. [38]. To homogenize the MR images across different sequences, skull stripping, co-registration and brain volume extraction was done. Then all the metastatic brain lesions were manually segmented using ITK snap toolbox. Multiple neural network architectures were fed with both unsegmented and manually segmented datasets. DL-based classification data was highly imbalanced as predominant sample size belongs to adenocarcinoma class and only a small proportion of cases belonged to squamous carcinoma groups. Hence the classification was based on the presence and absence of the EGFR mutation and ALK rearrangement. For the classification task, each class had less than 50 patients. To train the classifier, heavy data augmentation techniques like rotation, flip, etc. available in Keras were used. The augmented data was reshuffled for unbiased training. This dataset is then divided into 80% training and 20% test dataset. The models were trained to optimize the loss and to improve classification accuracy. DL model showed good predictive performance with pre-segmentation accuracy of 76% and post segmentation accuracy of 89%. The performance of the model in classifying the metastases improved significantly after the manual segmentation. During training and validation, since the case data was shuffled, class-wise true positive (TP), true negative (TN) and false positive (FP), false negative (FN) was not covered in the analysis. EfficientNetB0 with improved model scaling provides better accuracy over other CNN architectures and hence can be used to train complex data sets. No other published study shares similar objectives with this study.

This study concludes that both semantic features and DL model showed comparable accuracy in classifying EGFR mutation and ALK rearrangement in metastatic brain lesions. Application of both the methods in clinical practice may be useful to predict mutations in a patient while the biopsy or genetic mutation testing is awaited.

This was a site-specific study on NSCLC patients. Hence, the data is not applicable to other primary cancers such as breast or colon cancer. The radiomics and DL model were trained only on brain metastases from adenocarcinoma and a few cases of squamous cell carcinoma of the lung. Other genetic mutation i.e., ROS proto-oncogene 1, receptor tyrosine kinase (Ros-1) and Kristen rat sarcoma viral oncogene homolog (KRAS) mutations were not included in the current study. As each class had less than 50 patients, the augmented dataset was reshuffled for the classification task and for unbiased training of the classifier’s heavy data augmentation techniques. This dataset was then divided into 80% training dataset and 20% test dataset. Hence the class-wise analysis of true positive, true negative, false positive, and false negative was not feasible in this study. The performance of the proposed approach will significantly improve with a larger dataset in each class for more accurate classification.

In conclusion, both semantic features and DL models showed comparable accuracy in classifying EGFR mutation and ALK rearrangement. Both methods are non-invasive and can be clinically used to predict mutation status while invasive procedures-biopsy or genetic testing is undertaken. Application of both the methods in future clinical practice may help predict mutations in a patient while the biopsy or genetic mutation testing is awaited.