Title [Role Challenge - Origin Medical] AI Research Engineer.pdf ✅

1 Role Challenge - AI Research Engineer Bengaluru, Full time As a Research Engineer at Origin Medical, you will have the freedom to try and test hypotheses for multiple experiments as well as will be mentored to translate your research into a product for real-world impact. Here, you will learn real-time model development (TensorRT, TFLite, etc), working with huge datasets, while helping Origin Medical to develop solutions to real-world image processing problems. You will learn cloud computing (virtual machines, storage) based model training, logger for experiment tracking, and so on to advance your career path as a Research Engineer. Important Information and Evaluation Criteria ● Submission Details: a. Email your RC in a zip folder to [email protected] b. Email Subject: Research Team Role Challenge Submission. i. The submission folder should be named as Firstname_Lastname_Team eg. Siddhant Kumar for the Research Team should submit as Siddhant_Kumar_Research ii. Files required to be submitted are (MUST HAVES) 1. Model weights of all hypotheses tested 2. Trainer code (Python script/Jupyter Notebook) 3. Tester code (Python script/Jupyter Notebook) 4. Report (Format attached as Appendix-A) 5. ReadMe (Describing the details of files submitted) iii. Folder structure to be followed: iv. The framework preferred is PyTorch v. The code should be properly formatted according to PEP-8 guidelines. vi. Completing both Part A and Part B of the task is a must for evaluation. vii. The performance of the model will be evaluated on separate test images. viii. You may use GPU services such as Google Colab to train your models. ix. Plagiarism/Direct copying of online available codes are strongly discouraged. if found, will be disqualified straight away. x. Format and all files mentioned above for the role challenge are a must, failing which the role challenge will not be evaluated.
2 c. For queries contact: Ashrita Niharika Ashwin (Strategy Analyst) at [email protected] d. Please refer to the mail for the deadline for submission e. You are encouraged to submit earlier if you have completed the challenge instead of waiting till the deadline. Role Challenge When thinking about a career in AI, one most obvious question comes up: why and how? Well, it's not a straightforward question to answer but a few reasons we think why one should pursue a career in AI are: ● AI is called the skill of the century currently creating around 130 million roles across sectors including but not limited to medical diagnosis, autonomous vehicles, e-commerce, and even telecommunication. ● With advanced technologies, we have access to unlimited amounts of data. AI helps to derive useful information from these data and helps us understand the subject in more detail which could transform our insights. ● Application of AI ranges from image reconstruction of MRIs to preventive screenings in hospitals which impacts society in real time. ● AI career not only helps you gain great compensation but also gives a huge pool of opportunities right from computer vision to natural language processing and so on. Having said all the benefits and interests that come with AI, a lot of people struggle to get a head start. So more than the why question, the how question is more important. If you think you are suited for the role, the next step will be completing the role challenge. The role challenge aims to give you an interesting research hypothesis to solve offline to get a feel of what it's like to be a Research Engineer at Origin Medical. Task -To develop an algorithm that is capable of identifying the biparietal diameter (BPD) and occipitofrontal diameter (OFD) landmark points (2 per biometry) in fetal axial images. ● Ultrasound is the safest, most accessible, affordable, and widely used method for the detection of multiple fetal central nervous system (CNS) anomalies. ● Clinicians typically use biparietal diameter (BPD), occipitofrontal diameter (OFD), and head circumference to estimate gestational age and assess fetal growth and neurodevelopment, the first step in the detection of CNS anomalies. ● The figure shown below depicts (i) fetal brain ultrasound, (ii) ellipse fit for head circumference, and (iii) two biometry measurements (BPD and OFD can be considered as radii of the ellipse). ● Points A & C are two landmark points for the measurement of BPD and points b & d are two landmark points for the measurement of OFD.
3 ● Part A: Landmark Detection-Based Approach: ■ The dataset provided has 4 points (2 per biometry) for all the images. ■ The task here is to train a deep learning-based landmark detection model to predict the biometry points on the unseen images in the test set. ■ You can use this paper as a reference regarding the problem statement. It is optional to follow the same paper for the role challenge. 1. Regressing Heatmaps for Multiple Landmark Localization using CNNs 2. Cephalometric Landmark Detection by Attentive Feature Pyramid Fusion and Regression-Voting ■ Dataset link ● Part B: Segmentation-Based Approach i. The dataset provided has an ellipse fit for the cranium. ii. The task here is to train a segmentation model to identify the cranium and use a computer vision algorithm to find the 2 biometry points. iii. You can use this paper as a reference regarding the problem statement. It is optional to follow the same paper for the role challenge. 1. https://arxiv.org/pdf/1505.04597 2. https://arxiv.org/abs/1804.03999 iv. Dataset link ● Clinical Resources to understand the basics of fetal ultrasound: a. How To: Pregnancy BPD HC AC and FL Measurements 3D Video b. Transventricular View c. Performance of the routine mid-trimester fetal ultrasound scan ● What are we looking for? a. Your overall approach and the thought process to solve the problem - DO NOT worry about accuracy/performance metrics. b. Data pre-processing, data augmentations, etc., and the rationale for the same. c. What other models/modifications you tried to improve the performance of the tasks should be discussed here (At least 3 should be good). The hypothesis can be based on major changes in the pipeline, base architecture, pre-training or so. d. If given sufficient time, what are the methods you like to implement that you think will improve the performance. e. Key takeaways you will have from the results of the method you opted Most importantly, have fun while coding and enjoy doing the challenge! All the best!!! Please note: Keep Reading for the Appendix A - Report Format
4 Appendix A - Report Format 1. Motivation #Reason why the particular role challenge is selected 2. Abstract # Discuss in brief the method and results (150 words max) for each task attempted 3. Introduction # Brief discussion about the method opted, details on why and how the model/hypothesis is selected 4. Data PreProcessing/Analysis # All pre-processing tasks(Resizing, augmentation, etc.) and analysis (Ground truth check) should be discussed 5. Model Architecture #This section should answer questions but not be limited to - Why is the particular model selected? What is the benefit of the selected model in the current task? 6. Experimental Setting #Selection method of an experimental setup such as but not limited to optimizer, loss function 7. Hypothesis tried #What other models/modifications you tried to improve the performance of the tasks should be discussed here (At least 3 should be good). The hypothesis can be based on major changes in the pipeline, base architecture, pre-training or so. 8. Results #Critically analyzing the results obtained. 9. Key findings # Key takeaways you will have from the results of the best method you opted 10. Future work # If given sufficient time, what are the methods you like to implement that you think will improve the performance.