2019-04 MIA文章精选

• 2019 CVPR Oral Side Window Filtering, 龚元昊老师 [作者知乎解读] [code]

传统滤波方法不保边的原因是：都使用全窗口回归，会有沿着图像边缘的扩散。本文提出把窗口的边缘直接放在待处理像素的位置，这就切断了可能的法线方向的扩散。具体到一个像素位置，直接枚举八个可能的方向，让数据自适应地选择一个最佳的方向。

• 201903 Radiology 人工智能自动勾画鼻咽癌GTV，港中文Pheng-Ann Heng团队 [paper] [deepcare解读]

Deep Learning for Automated Contouring of Primary Tumor Volumes by MRI for Nasopharyngeal Carcinoma, 818训练，203测试；用20个测试数据比较AI和医生的分割结果。AI assistance improved contouring accuracy (five of eight oncologists had a higher median DSC after AI assistance; average median DSC, 0.74 vs 0.78; P < .001), reduced intra- and interobserver variation (by 36.4% and 54.5%, respectively), and reduced contouring time (by 39.4%). AI自动勾画然后医生修改，平均精度由74%提高至79%。

• 201904-腾讯深度解构产业互联网：九大领域打法，五个维度框架，[机器之心], [腾讯研究院介绍]

• 201904-前深度学习时代CTR预估模型的演化之路 [知乎王喆]

• MED NeurIPS 2018: Is your ML Methods solving a real clinical problem? by Tal Arbel

Focus lesion detection, segmentation, disease prediction in patient images
ML in Medical Imaging: patient diagnosis, understanding disease development, predicting patient outcome from images, personalized medicine.

ML方法没有被广泛应用到临床workflow的原因/挑战

• CV中的DL方法在医学图像中不总是work。比如BraTS分割任务DL很成功，但是存活时间预测任务效果不如人意。**Errors in performance lead to clinician mistrust.
• Clinicians don't trust black box methods. Interpretability is very important.
• No large scale annotated medical dataset for training. 导致通常在small, proprietary or benchmark dataset开发算法，缺乏鲁棒性。

Examine machine learning performance and metrics in real clinical contexts

• 临床影响：将病灶检测和分割算法加入商业软件中，提升了efficiency and precision，节省~5倍的时间和金钱；提升treatment analysis for almost all (22/23) new MS drugs in circulation wordwide. Clinical impact formula: Synergy with clinicians, end-users when designing method + trying methods and metrics for success to real clinical objectives = Clinical impact

201811-MICCAI 18 分割Decathlon冠军：3D Semi-Supervised Learning with Uncertainty-Aware Multi-View Co-Training，Nvidia. [arxiv]

Exploiting multi-viewpoint consistency for co-training.

LiTS测试集：95.9， 72.6

实验结果

• Why could a multiclass dice loss function solve the class imbalance problem?

In cross entropy, each pixel has the same weight irrespective of the class. by using a Dice loss, the weight of a pixel is different. If the CE tumor is small for example, then false positives or false negatives will impact the dice loss more and will thus intrinsically be weighted more.

• New roadmap outlines 5 research priorities for AI in radiology (radiology paper) (healthimaging报道)

1. Novel image reconstruction techniques that quickly produce images humans can read from source data.
2. A focus on automated image labeling and annotation, which includes “information extraction from the imaging report, electronic phenotyping and prospective structure image reporting.”
3. Machine learning models for clinical data, including pre-trained and distributed learning techniques.
4. Algorithms capable of explaining their findings to users.
5. Methods for deidentifying images and sharing image datasets that are adequately validated.

• FDA developing new rules for artificial intelligence in medicine (news)

• Can crowd-sourcing AI algorithms work in radiation oncology? (news) (JAMA paper)

• Data Science in one picture

• Segmentation models with pretrained backbones (pytorch)