[paper reading] Tun det: a novel network for meridian ultra sound nodule detection

[ Address of thesis ][ Code ][MICCAI 21]

Abstract

This paper presents a new single-stage detection model TUN-Det, Used for thyroid nodule detection in ultrasonic scanning . The main contribution is ：(i) introduce Residual U-blocks(RSU) To build our TUN-Det The backbone of ;(ii) Newly designed multi head architecture , By three parallel RSU Variant composition , To replace the ordinary convolution layer of classification and regression head . The remaining blocks enable local and global features to be extracted at each stage of the backbone , This plays an important role in detecting nodules of different sizes and appearances . Multi head design embeds the set strategy into an end-to-end module , Improve accuracy and robustness by fusing multiple outputs generated by different sub modules . stay 700 Of patients 1268 Experimental results on thyroid nodules showed , Our new proposal RSU Backbone and multi head structure for classification and regression head greatly improve the detection accuracy of baseline model . our TUN-Det In the overall average accuracy (AP) In terms of indicators, we have also achieved very competitive results compared with the most advanced models , And in $A{P}_{35}$ and $A{P}_{50}$ Better than them in terms of , This shows that it has good performance in clinical application .

Method

This paper belongs to a target detection task , To detect thyroid nodules from ultrasound images . The core idea is to use a more advanced Encoder Block(Residual U-blocks) To extract multi-scale features , As shown below ：

You can see , A major feature of this article is that it does not use any pre training backbone( Such as resnet) And so on. , Instead, it uses the Residual U-blocks(RSU). And this RSU It's actually a small one UNet, In the Encoder It improves the ability of multi-level feature extraction . Of course , The classification header and regression header in the article are also replaced by multi-layer In the form of , as follows ：

The implementation form is somewhat similar to that in the segmentation task deep supervision, That is to integrate the output of different levels for supervision , Instead of just monitoring the end result .

Generally speaking, it will U^2 Net[1] The idea of moving from saliency detection to target detection , In fact, the author of these two articles is the same .

Ref

[1] U2-Net: Going deeper with nested U-structure for salient object detection