TITLE

Meibomian Glands

WHO

Shixuan Li (sli221), Zhenhao Sun (zsun32), Mingzhi Zhang (mzhan146)

Introduction: What problem are you trying to solve and why?

In the final project, we are doing something new by implementing some native frameworks and image processing techniques. Our project is a task of amodal instance segmentation, which "amodal" means the capability to discover and reveal the overlapped parts in adjunct segments. We will be using the Meibomian Glands dataset as it is a perfect actual use case in helping diagnosis and the glands are densely distributed with a high percentage of overlapping. Traditional methods are not likely to work well in such cases (e.g, MaskR-CNN). The project can be used to provide reference for clinical diagnosis for dry-eye disease.

Related Work: Are you aware of any, or is there any prior work that you drew on to do your project?

Data: What data are you using (if any)?

We will be using the common meibomian glands dataset which contains a total of 1550 eyelid images, including upper-lid and lower-lid images. As for the upper-lid parts, due to the limited eyelid-flipping while taking the photos (images), we shall do image processing to manually extend the labeled areas so that it forms a complete actual gland (in original image, some parts of the glands are covered, so we would like to use image processing to fix it). The corrected images will be used as our dataset.

Methodology: What is the architecture of your model?

We shall be firstly try out some existing structures such as Mask R-CNN. Indeed, we already have done so, and the result is bad as expected, due to the overly dense and heavily overlapped glands distribution. Thus, the only way is to implement one-stage models. We selected DCAN, InstanceCut, and Discriminative Loss for experiments. We will try to modify the structures to see if it's possible to build an end-to-end amodal instance segmentation model. If not, we shall use two-stage model to achieve that (adding a generative model after our instance segmentation backbone). We already rented some GPUs for experiment.

Metrics: What constitutes “success?”

We plan to use precision, recall and mAP as the metrics for our instance segmentation task. Here recall is the True Positive Rate (for all the actual positives, how many are True positives predictions). Precision is the Positive prediction value (for all the positive predictions, how many are True positives predictions). The definition is proper since it indicates how well we are able to detect and finely segment our the glands. mAP stands for mean Average Precision where AP is defined as the area under the PR curve. We shall use Mask R-CNN as our baseline. In Mask R-CNN, the author uses AP50, AP75, AP-small, AP-median and AP-large for evaluation of COCO dataset (segmentation and bbox). We will not need the small/median/large metrics as the glands' shape are almost even. We will also not need the evaluation metrics for bounding boxes as we plan to use end-to-end models. Our goals is to bypass the baseline which is defined as Mask R-CNN's performance on our meibomian glands dataset.

Ethics: Choose 2 of the following bullet points to discuss; not all questions will be relevant to all projects so try to pick questions where there’s interesting engagement with your project. (Remember that there’s not necessarily an ethical/unethical binary; rather, we want to encourage you to think critically about your problem setup.)

Deep learning is a good approach for this task as DL is able to capture image features and give outputs based on image distribution. The dataset is all in gray-scale. Thus, it is very likely for clinics to miss some possible abnormal glands or crossed glands (unhealthy as glands should be located independently). Deep learning can provide very good reference or suggestions to help diagnosis. The dataset is a public dataset originated by UCSF. All privacy information has been removed, thus there is no social concerns. As the result of our model is likely to be used for doctor's reference, our result should be strong in recall than accuracy. This is because we are not particularly too strong on accuracy, as in real-life application, the model's main task is to provide suggestions on possible overlapped glands.

Division of labor: Briefly outline who will be responsible for which part(s) of the project.

Shixuan Li: Image processing (data cleaning, gland extension), Baseline preparation (Mask R-CNN). End-to-end methods training, testing and fine-tuning (Discriminative Loss). Zhenhao Sun: End-to-end methods training and testing(DCAN), Gland completion generative model (Pix2Pix) Mingzhi Zhang: End-to-end methods training and testing (InstanceCut). Proof and theory, visualizations (for all result data and intermediate auto-visualization system) and fine-tuning.

Built With

Share this project:

Updates