Name: 用於自動量化 HER2 擴增水準的細胞實例分割技術 (Cell Instance Segmentation for Automated quantifcation of HER2 amplification levels)
Uploaded: 2025-11-04T07:47:46.000Z
Duration: 4 min 57 s

Hi this is video demo for automated quantification of HER2amplification levels using deep learning.

這段影片示範如何使用 deep learning 自動量化 HER2擴增 的程度。

HER2 assessment is necessary for patient selection in anti HER2 targeted treatment.

HER2 評估對於選擇抗HER2 標靶治療的病患至關重要。

Current manual assessment methods are time costly, labor intensive, highly subjective and error prone.

目前的手動評估方法耗時、勞力密集、高度主觀且易出錯。

Visual challenges in fish and dish image interpretation such as unclear boundaries, shape variations and sparse labeling obstruct effective AI based analysis.

FISH 和 DISH 影像判讀中的視覺挑戰，像是界線不清、形狀變異和稀疏標記，都阻礙了有效的AI基礎分析。

We present an automated cascade deep learning approach and a signal detection model for precise HER2 and CEN17 quantification, assisting HER2amplification assessment to support targeted therapy selection in breast and gastric cancers through jacket based soft sampling loss.

我們提出一個自動級聯深度學習方法和一個訊號偵測模型，用於精確量化 HER2 和 CEN17，透過 jacket based soft sampling loss 協助 HER2擴增 評估，以支持乳癌和胃癌的標靶治療選擇。

Our model overcomes issues of overlapping cells, blurred boundaries and sparse labeling, ensuring balanced handling of complex cases and background noise which significantly boosts segmentation accuracy and model robustness.

我們的模型克服了細胞重疊、界線模糊和稀疏標記的問題，確保能平衡處理複雜情況和背景雜訊，顯著提升了分割準確度和模型穩健性。

We evaluated the proposed model with three clinical datasets of two different kinds of cancer.

我們使用三種不同癌症類型的臨床資料集來評估所提出的模型。

The data includes two clinical breast cancer datasets with 200 fish images and 210 dish images, each annotated by experienced pathologists as the reference standard.

這些資料包含兩個臨床乳癌資料集，共計 200 張 FISH 影像和 210 張 DISH 影像，每張都由經驗豐富的病理師標記作為參考標準。

Additionally, 45 dish images from gastric cancer cases are included.

此外，還包含 45 張來自胃癌病例的 DISH 影像。

Detailed information is provided in Table I.

This is main architecture of the proposed method.

A the FBN module with orange background color shows two concepts of top down sampling and skip connection to extract the intermediate convolutional feature Zeta using ResNext T50 as the backbone network.

帶有橘色背景的 FBN 模組，透過 ResNext T50 作為骨幹網路，展現了由上而下的取樣和跳接連接這兩個概念，以提取中間的卷積特徵 Zeta。

Then the features zeta are transferred into the RPN module to generate the bounding box prediction B1T and then moved into the B detection branch part with the green background color to generate the final bounding box result B B 1.

接著，特徵 zeta 會被傳送到 RPN 模組以產生邊界框預測 B1T，然後移至帶有綠色背景的 B 偵測分支部分，以產生最終的邊界框結果 B B 1。

R proposed soft sampling weighted loss function generates the new bounding box feature regression equations C Segmentation branch area with blue background color aims to pass the final prediction bounding boxes B the asterisk into the segmentation branch to generate the final prediction masks M D Instance segmentation results by the proposed method.

我們提出的軟取樣加權損失函數產生了新的邊界框特徵迴歸方程式。C 分割分支區域（藍色背景）旨在將最終預測的邊界框 B* 傳入分割分支，以產生最終預測的遮罩 M。D 執行個體分割結果由我們提出的方法達成。

FISH and dish technology allows for the simultaneous evaluation of the presence and localization of CEN17 and HER2 within cells determining HER2 gene amplification based on signal intensity and ratios.

FISH 和 DISH 技術能夠同時評估細胞內 CEN17 和 HER2 的存在與定位，並根據訊號強度和比例來判斷 HER2 基因擴增。

Color deconvolution is utilized to achieve color separation, allowing isolation and examination of specific components or features within an image for further analysis and interpretation.

顏色去卷積技術被用於實現顏色分離，從而分離和檢查影像中的特定組件或特徵，以便進一步分析和判讀。

A Raw FISH image sample B CN17 signal detection results highlighted in green using equation 31.

A 生 FISH 影像範例。B 使用方程式 31 以綠色標示的 CEN17 訊號偵測結果。

C results of signal detection of HER2 and partial CEN17 highlighted in yellow using equation 32.

C 使用方程式 32 以黃色標示的 HER2 和部分 CEN17 訊號偵測結果。

D detection results of CEN17, HER2 and HER2 related cells by the proposed method highl highlighted in green, red and blue respectively.

D 我們提出的方法偵測 CEN17、HER2 和 HER2 相關細胞的結果，分別以綠色、紅色和藍色高亮顯示。

B HER2 signal detection results highlighted in blue using equation 33C results of signal detection of CN17 and partial HER2 highlighted in pink using equation 33D detection results of HER2, CN17 and HER2 related cells by the proposed method highlighted in blue, pink and green, respectively.

B 使用方程式 33 以藍色標示的 HER2 訊號偵測結果。C 使用方程式 33 以粉紅色標示的 CN17 和部分 HER2 訊號偵測結果。D 我們提出的方法偵測 HER2、CN17 和 HER2 相關細胞的結果，分別以藍色、粉紅色和綠色高亮顯示。

Quantitative results the proposed cell instance segmentation model outperforms 10 baseline models in F1 score on the FISH breast cancer dataset and seven baseline models in F1 score on THE DISH breast cancer dataset while also showing promising predictive results on the DISH stomach cancer dataset.

量化結果我們提出的細胞個體分割模型在 FISH 乳癌資料集的 F1 分數上優於 10 個基準模型，在 DISH 乳癌資料集的 F1 分數上優於 7 個基準模型，同時在 DISH 胃癌資料集上也展現出有前景的預測結果。

Qualitative results Figure compares the segmentation outcomes of the proposed method with the benchmark approaches for HER2 related cells segmentation on FISH fluorescent microscope breast data.

質化結果圖比較了我們提出的方法與基準方法在 FISH 螢光顯微鏡乳癌資料上針對 HER2 相關細胞分割的結果。

Overall, the results show that there is a solid consistency between the instant segmentation results by the proposed method and the reference standard generated by the experienced pathologists.

總體而言，結果顯示我們提出的方法產生的即時分割結果與經驗豐富的病理學家所產生的參考標準之間有著穩固的一致性。

Results of the proposed method and reference standard on FISH fluorescent microscope breast data Figure provides a visual comparison of the results by the proposed method and the benchmark approaches for HER two related cell segmentation on dish light microscope breast data showing that the proposed method produces comparably promising results.

在 FISH 螢光顯微鏡乳癌資料上的我們提出的方法和參考標準的結果圖，視覺化比較了我們提出的方法與基準方法在 DISH 光學顯微鏡乳癌資料上針對 HER2 相關細胞分割的結果，顯示我們提出的方法產生了相當有前景的結果。

Results on Dish light microscope breast data Results on dish lite microscope stomach data thank you for your attention.

在 DISH 光學顯微鏡乳癌資料上的結果。在 DISH 光學顯微鏡胃癌資料上的結果。謝謝您的關注。

字幕列表影片播放

用於自動量化 HER2 擴增水準的細胞實例分割技術 (Cell Instance Segmentation for Automated quantifcation of HER2 amplification levels)

figure

determine

experience

prone