██FR█████ █INTELL███████████

Why One Small Parameter Can Make or Break Your Medical Image Retrieval


This content originally appeared on HackerNoon and was authored by Image Recognition

Abstract and 1. Introduction

  1. Materials and Methods

    2.1 Vector Database and Indexing

    2.2 Feature Extractors

    2.3 Dataset and Pre-processing

    2.4 Search and Retrieval

    2.5 Re-ranking retrieval and evaluation

  2. Evaluation and 3.1 Search and Retrieval

    3.2 Re-ranking

  3. Discussion

    4.1 Dataset and 4.2 Re-ranking

    4.3 Embeddings

    4.4 Volume-based, Region-based and Localized Retrieval and 4.5 Localization-ratio

  4. Conclusion, Acknowledgement, and References

4.4 Volume-based, Region-based and Localized Retrieval

Since multiple organs (i.e. labels) are present in each query volume, there are essentially different ways in which image retrieval can be performed. The preferred choice depends on the context of the retrieval task in the real world. If the goal is to find a scan out of a database that is most similar to a complete query scan with the entirety of all present organs (think scan-id to scan-id but visual), then volume-based retrieval is the right choice. In contrast, if the experimenter is interested in a particular organ and its most similar counterpart in the database (and all other organs just happen to be in the same scan due to proximity), then region-based retrieval or localized retrieval is advised. Slice-wise retrieval can find the most similar slice of a volume regardless of the number of other slices. This is not usually a practical choice in real scenarios. Figure 11 visualizes the options.

4.5 Localization-ratio

Table 22 shows the average localization-ratio for with and without re-ranking (L = 15). There is a drop for both 29 coarse and 104 original TS classes after re-ranking. However, both with and without re-ranking the localization-ratio is high, indicating that most of the slices contributing to the final retrieval of the volumes actually contain the desired anatomical region. However, based on (7) the choice of L can impact this measure. Figure 12 shows localization-ratio for different values of L as shown increasing the L to more that L = 15 decreases the localization-ratio. This indicates

\ Table 20: Summary of the average retrieval recall and standard deviation between classes for 29 anatomical regions, the boldfaced values highlight the highest recall across feature extractors.

\ Table 21: Summary of the average retrieval recall and standard deviation between classes for 104 anatomical regions, the boldfaced values highlight the highest recall across feature extractors.

\ Figure 11: An overview of three retrieval methods: (a) Slice-wise, where the retrieval is based on one selected slice e.g., the user zooms to a slice and retrieves the most similar slice (b) volume-based, where the retrieval is based on a complete volume, e.g., the user would like to retrieve similar volumes to the volumes under examination or simply filter the database (c) region-based, where the retrieval is based on the selected organ (or sub-volume), e.g, the user zooms in to a specific region and the most similar volume containing that region is retrieved, (d) localized, where the retrieval is based on the selected organ (or sub-volume) but the region in the retrieved volume is localized to the desired organ (or sub-volume), e.g, the user zooms in to a specific region and the system returns the localized region in the retrieved volume.

\ Figure 12: Average localization-ratio for different L values for re-ranking evaluation using DreamSim model (bestperforming embedding) based on (7).

\ Table 22: Summary of the average localization-ratio and standard deviation between classes for 29 and 104 anatomical regions for L = 15, the boldfaced values highlight the highest localization-ratio across feature extractors.

\ that the first highest values in vector mSIM based on (6) points to the exact slices that contain the desired anatomical structure. To improve the re-ranking localization-ratio, further studies can focus on optimizing the L value based on heuristic, organ size, etc.

\

:::info Authors:

(1) Farnaz Khun Jush, Bayer AG, Berlin, Germany (farnaz.khunjush@bayer.com);

(2) Steffen Vogler, Bayer AG, Berlin, Germany (steffen.vogler@bayer.com);

(3) Tuan Truong, Bayer AG, Berlin, Germany (tuan.truong@bayer.com);

(4) Matthias Lenga, Bayer AG, Berlin, Germany (matthias.lenga@bayer.com).

:::

:::info This paper is available on arxiv under CC BY 4.0 DEED license.

:::

\


This content originally appeared on HackerNoon and was authored by Image Recognition