Screening strategies for cancer, the second largest cause of deaths, exist, but are invasive, cumbersome, and expensive. Many cancers lack viable screening modalities all together. Circulating tumor cell clusters (CTCCs) are seen during early stages of cancer and are larger than normal blood cells. Discrimination of such differential sizes by real-time ultrasound scanning of a blood vessel offers an attractive universal screening tool for cancer. Yeast colonies were grown to different sizes mimicking CTCCs and normal blood cells, using sugar and starch to incubate and sodium fluoride to arrest growth after specified times. They were circulated using syringes and an infusion pump through a wall-less ultrasound phantom, made using agar (mimicking human soft tissue), and Doppler ultrasound was performed, with screenshots taken. Key characteristics of particles of interest were identified. Ultrasound data were processed and used to train a convolutional neural network (CNN). Six models with binary classification were tested. Doppler signals of CTCC surrogates could be visually distinguished from normal cells and normal saline, proving the principle of ultrasound size discrimination of CTCCs. The most accurate machine learning model yielded 98.35% accuracy in the prediction of CTCCs, exceeding human evaluation accuracy. Thus, machine learning could help automate and improve detection of cancer by screening for CTCCs.

癌症是第二大死因，其筛查策略是存在的，但具有侵入性、繁琐且昂贵。许多癌症缺乏可行的筛查方式。循环肿瘤细胞簇 (CTCC) 在癌症早期阶段可见，比正常血细胞大。通过对血管进行实时超声扫描来区分这种差异尺寸，为癌症提供了一种有吸引力的通用筛查工具。酵母菌落生长成不同大小，模仿 CTCC 和正常血细胞，使用糖和淀粉进行孵育，并使用氟化钠在指定时间后阻止生长。使用注射器和输液泵通过使用琼脂（模仿人体软组织）制成的无壁超声模型进行循环，并进行多普勒超声检查并截取屏幕截图。确定了感兴趣颗粒的关键特征。超声数据经过处理并用于训练卷积神经网络 (CNN)。测试了六种二元分类模型。CTCC替代物的多普勒信号可以在视觉上与正常细胞和生理盐水区分开，证明了超声区分CTCC大小的原理。最准确的机器学习模型对 CTCC 的预测准确率高达 98.35%，超过了人类评估的准确率。因此，机器学习可以通过筛查 CTCC 来帮助自动化和改进癌症检测。