【USE】《An End-to-End System for Automatic Urinary Particle Recognition with CNN》
Urine Sediment Examination(USE)
JMOS-2018
目錄
- 目錄
- 1 Background and Motivation
- 2 Innovation
- 3 Advantages
- 4 Methods(Meta-architectures)
- 5 Experiments
- 6 Conclusion
1 Background and Motivation
The urine sediment analysis of particles in microscopic images can assist physicians in evaluating patients with renal
and urinary tract diseases. Manual urine sediment examination
The Overfeat made the earliest efforts to apply deep CNNs to learn highly discriminative yet invariant feature for object detection. 作者採用基於 deep learning 採用 CNN 提取特徵的方式,來取代 hand-crafted features
傳統的 multi-stage 方法 heavily depends on the accuracy of the segmentation and the effectiveness of the hand-crafted features.
基於CNN 的方法 可以可以實現 end-to-end,segmentation free and 提取的 feature more discriminatory
2 Innovation
- Exploit Faster R-CNN and SSD for urine particle recognition
- Investigate various factors to improve the performance of Faster R-CNN and its variants
- Trimmed SSD to achieve better performance
3 Advantages
- Best mAP of 84.1%(準),get a best AP of 77.2% for cast particles
- Only 72ms per image for 7 categories(快)
4 Methods(Meta-architectures)
- MS-FRCNN(multiple scale Faster RCNN)
- OHEM-FRCNN(Faster R-CNN with online hard example mining)
- Trimmed SSD
詳細的說明
Faster RCNN :shareable CNN feature extraction + region proposal generation + region classification and regression,採用 a pyramid of anchors
MS-Faster RCNN:builds a more sophisticated network for Fast R-CNN detector by a combination of both global context and local appearance features.
OHEM-Faster RCNN:Instead of a sampled mini-batch, it eliminates several heuristics and hyperparameters in common use and selects automatically hard examples by loss.
SSD:Unlike YOLO, it improves detection quality by applying a set of small convolutional filters to multiple feature maps to predict confidences and boxes offsets for various-size categories.
Trimmed SSD: 作者的資料集類別數比較少,SSD直接拿來用,會 produce a large number of redundant prediction results interfering with the final detection performance. For simplification, we attempt to remove several top convolutional layers from the auxiliary network of SSD, which leads to the trimmed SSD.
removing conv7, conv8, and conv9 layers
5 Experiments
5.1 datasets
Dataset consisting of 5,376 annotated images corresponding to 7 categories:
- erythrocyte (紅細胞)目標數:21,815
- leukocyte(白細胞)目標數:6,169
- epithelial cell(上皮細胞)目標數:6,175
- crystal(結晶)目標數:1,644
- cast(管型)目標數:3,663
- mycete(黴菌)目標數:2,083
- epithelial nuclei(上皮核)目標數:687
資料集分佈情況
5.2 Trainning
5.2.1 Feature extractors
(ZF、VGG、ResNet-50、ResNet-101、PVANet)
5.2.2 Training strategies
- 4 steps as Faster RCNN
- approximate joint training(end-to-end training)
end-to-end 比較好
5.3 不同scales 和 backbones 比較
不同 backbone,anchor 的不同 scales(ratios都是 1:1,1:2,2:1,因為資料集的 object 比較小,所以增加了scales的種類) 的結果如下,PVANet 比較好。
5.4 Data augmentation
a horizontal flip to augment training set
下圖展示了 horizontal 和 verticle flip 的 比較,單獨用都有提升,一起用沒有提升,一般都是用 horizontal,為啥 vertical 也會有提升呢,個人感覺因為資料集是細胞,所以形狀在豎直方向翻轉,影響沒有那麼大。
5.5 Faster RCNN vs MS-Faster RCNN
從表格可以分析,MS-Faster RCNN 的效果會比 Faster RCNN 差,但是隨著 anchor scales 的 diversity 增加,他們的之間的 gap 會縮小,且 MS-Faster RCNN 在小目標上會有更好的效果。
5.6 Faster RCNN vs Faster RCNN+OHEM
加了效果好,資料集越多,more benefits
5.7 SSD vs Trimmed SSD
為了適應小目標,smaller is better
5.8 Adding bells & whistles
5.8.1 anchor scales
- the more the better
- the smaller the superior
下圖(a)VGG-16 為例,不同 anchor scales 的 proposal recall,(b)是不同 backbones 的 proposal recall,(c)不同 backbones 的 mAP
5.8.2 Feature extractors
圖6 (b),用不同的 backbones
5.8.3 PVANet vs. VGG-16
由圖6(c)可以看出,PVANet的 proposal 質量會差(曲線下降的比較快),但是由 table 2 看出,他最終的結果比較好,下圖是檢測時 recall 和prediction 的圖,可以看出,隨著 recall 增加,PVANet 的 precision 相較於 VGG-16 下降的更慢,且比 VGG-16 高。
6 Conclusion
在Faster RCNN 和 SSD 的基礎上結合自己的資料集,用不同的 backbones,anchor scales,training stages 來提升 mAP。
- MS Faster RCNN
- Trimmed SSD(去掉一些層)