2
TITLE: Interactive medical 3D segmentation using deep learning and model pruning GUIDANCE: Adriaan Van Gerven; [email protected]; relu BV; kappeldreef 60 3001 Heverlee PROMOTOR: Prof. Tinne Tuytelaars DAILY SUPERVISORS: Adriaan Van Gerven; Thomas Beznik NUMBER OF STUDENTS: 1 CONTEXT: In Dentistry, Cone Beam Computed Tomography scans (CBCT-scans) are increasingly being used to diagnose and to plan various treatments. A crucial step for this is to have an accurate segmentation of each tooth in the scan. Relu has created AI-pipelines to automatically segment all the teeth inside of a CBCT. However a 100% perfect output can never be guaranteed, therefore the user needs to be able to edit the segmentations provided by the AI. For this Relu would like to develop another AI which can take user input such as clicks and strokes and adapts the segmented shape accordingly. Before these techniques can be used in production, several challenges still are unsolved such as: hard constraints on the edits, real-time performance & memory efficiency on the 3D regions. Besides the interactive segmentation techniques the student should therefore also research model pruning techniques to demonstrate the feasibility of this interactive segmentation in a production context. Depending on the students interest and the timing the scope of the project can be extended to other maxillofacial structures. The company will provide the data and a pipeline that is ready to use as well as an interface to test the implemented AI architectures so that the student can focus on learning, implementing and testing various AI techniques. COMPANY: Relu is a startup conveniently located on campus Arenberg. Relu aims to radically improve the way that oro-maxillo-facial (teeth and skull) treatments are done by leveraging the power of AI behind an intuitive user interface. The students will directly be coached by the founding team and will be immersed in an entrepreneurial environment focused on AI. GOAL: The goal is to research, implement and evaluate various techniques for 3D interactive segmentation and model pruning to bring this tool to production. The student can come up with his own ideas of architecture and is given the freedom to experiment. METHODOLOGY: The student will perform the following steps: Understand the provided data Research the current state-of-the art in 3D interactive segmentation & model pruning Get a baseline model working and implement existing architectures or new approaches Perform model pruning Evaluate and report on the trained model PROFILE/REQUIRED SKILLS: Students are expected to know python or have a strong drive to learn it fast. Experience in

GUIDANCE: Adriaan Van Gerven; · 2020. 10. 2. · TITLE: Interactive medical 3D segmentation using deep learning and model pruning GUIDANCE: Adriaan Van Gerven; [email protected];

  • Upload
    others

  • View
    4

  • Download
    0

Embed Size (px)

Citation preview

Page 1: GUIDANCE: Adriaan Van Gerven; · 2020. 10. 2. · TITLE: Interactive medical 3D segmentation using deep learning and model pruning GUIDANCE: Adriaan Van Gerven; adriaan.vangerven@relu.eu;

TITLE: Interactive medical 3D segmentation using deep learning and model pruning GUIDANCE: Adriaan Van Gerven;

[email protected]; relu BV; kappeldreef 60 3001 Heverlee

PROMOTOR: Prof. Tinne Tuytelaars DAILY SUPERVISORS: Adriaan Van Gerven; Thomas Beznik NUMBER OF STUDENTS: 1 CONTEXT: In Dentistry, Cone Beam Computed Tomography scans (CBCT-scans) are increasingly being used to diagnose and to plan various treatments. A crucial step for this is to have an accurate segmentation of each tooth in the scan. Relu has created AI-pipelines to automatically segment all the teeth inside of a CBCT. However a 100% perfect output can never be guaranteed, therefore the user needs to be able to edit the segmentations provided by the AI. For this Relu would like to develop another AI which can take user input such as clicks and strokes and adapts the segmented shape accordingly. Before these techniques can be used in production, several challenges still are unsolved such as: hard constraints on the edits, real-time performance & memory efficiency on the 3D regions. Besides the interactive segmentation techniques the student should therefore also research model pruning techniques to demonstrate the feasibility of this interactive segmentation in a production context. Depending on the students interest and the timing the scope of the project can be extended to other maxillofacial structures. The company will provide the data and a pipeline that is ready to use as well as an interface to test the implemented AI architectures so that the student can focus on learning, implementing and testing various AI techniques. COMPANY: Relu is a startup conveniently located on campus Arenberg. Relu aims to radically improve the way that oro-maxillo-facial (teeth and skull) treatments are done by leveraging the power of AI behind an intuitive user interface. The students will directly be coached by the founding team and will be immersed in an entrepreneurial environment focused on AI. GOAL: The goal is to research, implement and evaluate various techniques for 3D interactive segmentation and model pruning to bring this tool to production. The student can come up with his own ideas of architecture and is given the freedom to experiment. METHODOLOGY: The student will perform the following steps:

● Understand the provided data ● Research the current state-of-the art in 3D interactive segmentation & model pruning ● Get a baseline model working and implement existing architectures or new

approaches ● Perform model pruning ● Evaluate and report on the trained model

PROFILE/REQUIRED SKILLS: Students are expected to know python or have a strong drive to learn it fast. Experience in

Page 2: GUIDANCE: Adriaan Van Gerven; · 2020. 10. 2. · TITLE: Interactive medical 3D segmentation using deep learning and model pruning GUIDANCE: Adriaan Van Gerven; adriaan.vangerven@relu.eu;

pytorch and/or a (bio)medical background is a plus. A strong mathematical background is required to make this internship a success. COVID-19: Students can continue their internship online in case of governmental restriction. Work in the office is preferred.

Figure 1 Examples of segmentation errors to be corrected