## 材料和方法

### 用深度学习的方法分割细胞核

Overview of the nuclei segmentation procedure. $DRAN_{BL}$ and $DRAN_{BD}$ are the models for nuclei blob detection and boundary detection, respectively.

#### 多尺度聚合

Multiscale Deep Residual Network (MDRAN) architecture. MDRAN composes of 3 DRANs at 3 scales (x0.5, x1.0, x2.0) and a decoder (in dash rectangle), aggregating 3 scales together and generating a segmentation map at x1.0 scale. In the decoder, the convolution block [128, 5 × 5, 256] denotes [128 input channels, 5 × 5 kernel, 256 output channels].

### 两步的自动化方法对整个载片图像进行分类

Overview of the NSCLC classification framework. (A) Workflow for training the neural network to classify input patches as either non-diagnostic (ND), lung adenocarcinoma (LUAD), or lung squamous cell carcinoma (LUSC). (B) Workflow for processing the WSIs within the test set to obtain probability maps for each class. (C) Workflow for the random forest regression model. Features are extracted from LUAD and LUSC probability maps and then fed as input into the random forest model. SN stands for stain normalization by method of Reinhard et al. (2001).

## 结果

### 分割效果评估方法

DICE系数衡量了实际上和算法输出之间的重叠部分，但是没有考虑到分离和融合的情况，“分离”是本是一个核，算法输出是多个核，“融合”是本是多个核，算法输出一个核。使用DICE系数，一个算法把两个接触或者重叠的核看作一个和正确的分割为两个，具有相同的DICE。DICE-2可以解决“分割”的问题。伪代码如下：

Q and P are the sets of segmented objects (nuclei). The two DICE coefficients were computed for each image tile in the test dataset. The score for the image tile was calculated as the average of the two dice coefficients. The score for the entire test dataset was computed as the average of the scores of all the image tiles.

### 分割细胞核的深度学习方法实验性评估

1.考虑到补丁的宽度和高度，在[-0.05,0.05]范围内的随机垂直和水平位移
2.在[-45°，45°]度范围内随机旋转
3.随机垂直和水平翻转，概率为0.5
4.随机剪切，强度范围为[-0.4π，0.4π]
5.随机调整大小，比率范围为[0.6,2.0]

Image patch generation. To avoid zero-padding in augmentation, a patch of size 200 × 200 is first provided. Subsequently, the center region of 102 × 102 is cropped and fed into the network as input. For an input of size 102 × 102, the network provides a segmentation map of size 54 × 54.

Head-to-head comparison between MDRANBL and DRANBL on the test set. Test images are ordered by the ascending order of MDRAN DICE_1. The shaded area indicates that the images were scanned at 20x magnification.

Examples of nuclei segmentation via the multiscale aggregation (MDRANBL+DRANBD) and single scale (DRANBL+DRANBD) approach. The images from top to bottom are the 1st, 11th, 20th, and 25th image tile in the test set.

Examples of correct and incorrect nuclei segmentation. Our method (Bottom) is able to distinguish the boundary of the non-highly overlapping nuclei fairly well but (Top) fails on the highly overlapping nuclei with disproportionate stains.

### 整个载玻片图像的分类

Test WSIs with overlaid probability maps. Blue/purple indicates a region classified as diagnostic LUAD, green indicates a region classified as diagnostic LUSC, yellow/orange refers to a region classified as non-diagnostic.