Lipid droplet detection in histopathological images using reinforcement learning

Figures & data

Figure 1. Flowchart describing the proposed method.

Figure 2. Weights of an error diffusion filter (*target pixel).

Figure 5. ‘AND’ image (logical product between two images) of each rank: rank 6 (a), rank 5 (b), rank 4 (c) and rank 3 (d) [scale bar: 100 μm]. *A different version of the image in Figure 3(a) is displayed.

Figure 6. Environment: numbers indicate indexes, circle indicates an agent, and arrow indicates movement direction.

Figure 7. Definition of reward: overlapping region/newly extracted region.

Figure 8. Episodes and average Q values.

Figure 3. Original image 1/6 (a) and cell nucleus image (b) [scale bar: 100 μm]. *1/6 of the original image in Figure 3(a) is displayed.

Figure 4. Foreground binary image of each rank: rank 6 (a), rank 5 (b), rank 4 (c), and rank 3 (d).[scale bar: 100 μm]. *A different version of the image in in Figure 3(a) is displayed. Four types of foreground binary images were created, each corresponding to a specific rank (6, 5, 4 and 3) as a threshold. The figure illustrates the foreground binary images for each rank in the original image (1/6 area).

Figure 9. Each rank image with noise removed: rank 6 (a), rank 5 (b), rank 4 (c) and rank 3 (d). [scale bar: 100 μm]. *A different version of the image in Figure 3(a) is displayed.

Figure 10. Lipid droplet detection image: each rank (a)–(d) and overall (e) [scale bar: 100 μm]. *A different version of the image in Figure 3(a) is displayed.

Table 1. Classifier comparison.

Classifier	F-value
k-nearest neighbour method	0.8812
SVM	0.9013
RF	0.9016

Table 2. Effect of dimensionality reduction.

Evaluation index	RF	Dimensional reduction RF
Accuracy	0.89	0.855
Precision	0.868	0.833
Recall	0.92	0.90
f-value	0.893	0.862

Figure 11. Scatter plot matrix.

Table 3. Correct answer rate for two lipid droplet images [%].

Rank	Image A	Image B
6	–*	82.4 (14/17)
5	100.0 (1/1)	90.9 (40/44)
4	83.2 (109/131)	75.3 (207/275)
3	90.9 (748/823)	64.9 (378/582)

* Rank 6 of lipid droplet image A is not detected.

Figure 12. Original image 1/6 (a) and lipid droplet detection image (b). [scale bar: 100 μm]. *A different version of the image in Figure 3(a) is displayed. The detection performance was relatively effective for large lipid droplets, but smaller lipid droplets were identified with lower accuracy, resulting in detection gaps. The execution time for this method averaged approximately 16 min and 30 s per image using Google Colaboratory-1.0.0.

Data availability statement

The data supporting the findings reported herein are available from the corresponding author, HN, upon reasonable request.