2025 年 1 月 17 日，65 岁的丘先生因头晕三天走进宁波大学附属人民医院。头晕，在老年人身上是再寻常不过的症状，可能因为脑供血不足、颈椎问题或血压波动。医生按照惯例安排了彩超和平扫 CT，结果没有明显异常，丘先生本应拿着报告回家休养。然而，AI 发出了警报。
系统扫描了那张平扫 CT，在胰腺区域标记出一处直径约 1.5 厘米的病灶，提示恶性病变可能。主治医生愣了一下，丘先生的症状是头晕，没有任何指向胰腺的迹象。
头晕三天，一场与“癌王”的意外相遇
宁波大学附属人民医院副院长、肝胆胰普外科主任医师朱柯磊后来回忆：“这是我们医生都想不到的，因为患者的主诉是头晕，而胰腺有问题一般是腹痛。”
AI 的报警让医院迅速调整方向，增强 CT 随即安排，结果明确指向胰腺癌。丘先生很快接受了手术，病理报告证实为胰腺导管腺癌。他的女儿说：“太庆幸了，我父亲真是捡回一条命。”
胰腺癌素有“癌王”之称。胰腺深埋于腹腔，被胃、十二指肠、脾脏团团包围，位置隐蔽，胰腺癌早期几乎没有特异性症状。等到患者感受到明显不适，癌细胞往往已经悄然转移。
中国每年新发胰腺癌患者超过 13 万人，超过 80% 确诊即晚期，五年生存率不到 10%。
用朱柯磊的话说，AI 给了医生一双特殊的眼睛，从完全不同的角度来审视病情。
2024 年 11 月起，宁波大学附属人民医院引入阿里达摩院（湖畔实验室）的医疗 AI，联合上海市胰腺疾病研究所金钢教授团队开展前瞻性临床试验，成为浙江省内率先开设胰腺癌早筛门诊的医院。
截至目前，宁波项目已从 4 万余人的 CT 影像数据中发现 8 例胰腺癌，其中 6 例为早期，全部接受了手术治疗。
达摩院的 PANDA 模型研究团队将 2000 余例胰腺癌患者的增强 CT 影像配准到平扫 CT 上，让模型学会在分辨率更低的图像中识别胰腺的细微病变。
模型对胰腺癌患者的识别敏感性达 92.9%，特异性达 99.9%，比有经验的影像科医生高出约 34% 的敏感性。
增强 CT 需要注射造影剂、费用较高，不适合大规模人群筛查；而平扫 CT 是门诊中最普遍的影像检查手段，如果能在这些现成的影像里完成癌症的“机会性筛查”，意味着不需要给患者额外增加任何检查负担，任何一次普通就诊或体检都可能成为发现早期癌症的机会窗口。
达摩院多癌早筛 AI 技术已开始全球推广，覆盖全球 9 个国家和地区，服务超过 2000 万人次。
与此同时，达摩院也在推进“一扫多筛”的目标，即通过一次平扫 CT，AI 不仅识别胰腺癌，还能同步完成肝癌、食管癌、胃癌、结直肠癌乃至主动脉综合征等急症的筛查。
继胰腺癌之后，达摩院发布的肠癌筛查 AI 模型 DAMO COCA，从 2.7 万人的平扫 CT 影像中精准识别出 5 例漏诊肠癌，敏感性达到 86.6%、特异性达到 99.8%。
医生在经验判断和临床决策上无可替代，但人眼存在疲劳盲区，注意力会被主诉牵引，面对几百张细节密集的影像切片，总有遗漏的可能。
AI 没有疲劳，不会被主诉干扰，不存在“今天看太多片子了”的状态波动，因而发现了那个原本可能被推迟到晚期才露面的病灶。
AI 已在用你的脸，预判你的命运
AI 在癌症领域的探索并不止于此。
近日，哈佛大学麻省总医院的研究团队在《Lancet Digital Health》上发表了一项研究，他们用一张普通照片，预测一个癌症患者的死亡风险。虽然该模型的预测结果并非完全准确，但那些死亡风险较高的患者，其死亡概率确实更高。
FaceAge 的逻辑建立在“生物年龄”的概念上。同样 60 岁，有人看起来像 52 岁，有人看起来像 71 岁，是因为生活方式、疾病史和遗传上的差异。
研究员 Mak 指出，吸烟的人往往看起来比实际年龄要老一些；酗酒的人也是如此。
研究者们训练了一种算法，使其能够在不给定特定任务时，分析和处理 4000 万张普通人脸图像。随后，他们调整了该模型，使其能够以两种不同方式分析人脸图像：一种用于估算一个人的生理年龄，一种用于预测癌症患者的生存几率。
AI 倾向于基于面部某些区域来预测生理年龄，比如鼻唇沟（我们也称之为上巴），以及前额的部分区域。而在死亡风险预测方面，AI 则倾向于关注鼻梁和眼睛下方的区域。
无论照片中的人是在皱眉还是微笑，无论是否化妆以及处于何种光线环境下，这些都不会对年龄预测产生影响。
AI 还能够考虑到常见的整形手术情况。不过，接受鼻整形手术的人往往看起来比手术前的照片年龄稍大一些。
死亡风险预测方面，在一组由 3820 名癌症患者组成的测试组中，死亡风险评分最高的那一组中超过半数患者在拍照后短短 90 天内便去世了。
研究者还在肺癌患者的基因数据里发现，FaceAge 的评分和一个叫 CDK6 的基因呈负相关。CDK6 是细胞周期调控的重要参与者，与多种癌症的进展有关。一个靠脸部照片训练出来的模型，竟然在某种程度上捕捉到了基因层面的信号。
Sophont 的首席执行官在社交媒体上分享了另一个案例，涉及骨肉瘤。
他们的模型 Holon-S1 分析了患者的病理切片。这个模型没有经过骨肉瘤的专项训练，但它从切片图像里提取出了关键的生物学信息，患者的成纤维细胞相关基因出现了过度表达现象。
这一现象之前已经由该患者的医疗团队通过先进的分子检测手段发现，并且已经确定了相应的治疗靶点。
新的问题
FaceAge 的研究者在论文里专门提到了一个让他们忧虑的方向：这样的工具对临床医生而言可能是一种助力，但同时也可能对寿险公司或雇主极具吸引力，他们或许会尝试借此来评估个人的风险水平。
丘先生被 AI 从漫无目的的就诊人流中“拎出来”的那一刻，很幸运，是技术的善意应用。而未来某个保险申请人因为一张照片被 AI 标记为高风险，则带来了新的问题。
此外，模型学到的“衰老的视觉特征”，很可能并不适用于所有肤色和面部结构。如果模型对某些族裔系统性地高估生物年龄，这个误差会以一种极其隐蔽的方式，复刻并放大已经存在的不平等，而被误判的人甚至不知道发生了什么。
曾有研究发现，某些用于预测患者医疗需求的商业模型，对黑人患者的需求存在系统性低估，因为模型用历史医疗花费作为健康状况的代理指标，而历史上黑人患者在相同病情下所获得的医疗资源本就更少。
研究者强调，FaceAge 在进入临床之前，需要在更多样化的人群中测试，识别并消除潜在的误差。
如何处理这些新的问题，比训练一个模型难得多。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On 17 January 2025, Mr. Chu, aged 65, entered the People ' s Hospital attached to Ningbo University for three days because he was dizzy. Dizziness is an unusual symptom in older persons, which may be due to insufficient brain blood, cervical problems or blood pressure fluctuations. As a result of the usual practice of CT, which had been arranged by doctors, there was no apparent anomaly, and Mr. Chu should have taken the report home for rest. However, AI issued an alarm.
The system scanned the flat CT and marked a stove about 1.5 centimetres in diameter in the pancreas area, suggesting the possibility of a malignant disease. The attending doctor paused, and Mr. Chu was suffering from dizziness and no sign of pointing to pancreas.
Three days dizzy, an accident with the Cancer King.
The Vice-President of the People's Hospital attached to the University of Ningbo, Dr. Zhu Koo-Pil, later recalled: “This was unexpected to our doctors, because the patient's main complaint was dizziness, while the pancreas problems were usually abdominal pain”.
The alarm in AI allowed the hospital to re-route quickly and enhance the CT arrangements immediately, with the result clearly pointing to pancreas cancer. Mr. Chu was immediately operated on, and the pathology report confirmed cancer of the pancreatic catheter. His daughter said, "I'm so glad my father got his life back."
Pancreas cancer is called Cancer King. The pancreas are buried deep in the abdominal cavity, surrounded by stomachs, gestals and spleen, and their location is hidden, and early pancreas cancers are almost non-specific. By the time the patient feels clearly ill, cancer cells tend to shift.
More than 130,000 new cases of pancreas cancer occur in China each year, more than 80 per cent of which are diagnosed at an advanced stage and less than 10 per cent in five years.
In the words of Zhukilian, AI gave the doctor a special eye and looked at the situation from a completely different angle.
Since November 2024, the People ' s Hospital attached to Ningbo University has introduced the Aridamo Hospital (Landside Laboratories) in AI, a forward-looking clinical trial conducted by a team of Golden Steel Professors at the Shanghai Pancreas Disease Institute, which became the first hospital to open an early screening clinic for pancreas cancer in Zhejiang province.
To date, the Ningbo project has detected 8 cases of pancreas cancer from CT images of over 40,000 people, of which 6 were early and all were treated for surgery.
The PANDA Modelling Team at the Dama House has calibrated an enhanced CT image of 2000 cases of pancreas cancer to flatten the CT, so that the model learns to recognize the nuclei of pancreas in a lower resolution image.
The model has a recognition sensitivity of 92.9 per cent to pancreatic cancer and 99.9 per cent to about 34 per cent more than an experienced video doctor.
The enhancement of CT requires a high cost of being injected with a film maker, which is not suitable for large-scale mass screening; and CT is the most common means of video screening in clinics. If the “opportunity screening” of cancer can be completed in these ready-made images, it means that no additional examination burden is required for the patient, and any regular visit or medical examination may be a window of opportunity to detect early cancers.
The Al technology for cancer early screening at the Dharma complex has begun to be rolled out globally, covering nine countries and territories around the world, serving more than 20 million people.
At the same time, Dharma House is also pursuing the goal of “sifting multiple sift” through a clean-up of CT, AI not only identifying pancreas cancer, but also synchronizing the screening of acute diseases such as liver, oesophagus, stomach, colon and even arctic syndrome.
Following pancreas cancer, DAMO COCA, an AI model of publish Dharma Hospital, identified with precision 5 cases of out-of-body CT cancer, with a sensitivity of 86.6 per cent and a specificity of 99.8 per cent, from a clean-up of 27 million CT images.
There is no substitute for doctors in empirical judgement and clinical decision-making, but there are tired and blind spots in which attention can be drawn by the lead, and there is always the possibility of missing out in the face of hundreds of detailed video slices.
AI was not tired, was not disturbed by the main suit, and there were no fluctuations in the status of “too many movies today”, which led to the discovery of a stove that might have been delayed until late.
AI is already using your face to predict your fate.
AI explores more than that in the field of cancer.
In recent days, the research team at the Massachusetts General Hospital at Harvard University published a study in Lancet Digital Health, which predicts the risk of death of a cancer patient using a common photograph. Although the model ' s projections are not entirely accurate, those at higher risk of death are more likely to die.
FaceAge's logic is based on the concept of “bioage”. Also 60 years old, someone looks 52 years old, someone looks 71 years old, because of differences in lifestyle, disease history and genetics.
Researcher Mak noted that smokers tend to look older than they actually are; the same applies to alcoholics.
Researchers trained an algorithm that enabled them to analyse and process 4 million images of ordinary human faces without giving a specific assignment. They then adjusted the model to enable them to analyse the image of the human face in two different ways: one for estimating a person ' s biological age and the other for predicting the chances of survival of cancer patients.
AI tends to predict biological age based on certain areas of the face, such as the nasal lasagna (which we also call the upper bar) and parts of the forehead. In terms of mortality risk projections, AI tends to focus on the nostrils and areas below the eye.
Whether the people in the photographs frown or smile, whether or not they wear makeup and in whatever light, these do not have an impact on age projections.
AI is also able to take into account common plastic surgery. However, those undergoing nasal prosthesis often appear to be slightly older than the photographs before.
In terms of mortality risk predictions, in a group of 3820 cancer patients, more than half of the group with the highest death risk rating died within 90 days of taking the photo.
Researchers also found in genetic data for lung cancer patients that FaceAge's rating was negative for a gene called CDK6. CDK6 is an important participant in cytological cyclical regulation, linked to advances in multiple cancers. A model trained by face photographs captures, to some extent, signals at the genetic level.
The CEO of Sophont shared another case in social media, involving bone tumours.
Their model Holon-S1 analyzes the pathological slices of the patient. The model is not specially trained for osteoporosis, but it extracts critical biological information from the sliced images and overexpression of patients' musculos fibre cell-related genes.
This phenomenon was previously detected by the patient ' s medical team through advanced molecular tests and the corresponding targets for treatment have been identified.
New issues
FaceAge's researchers in their paper singled out a direction that worries them: Such tools may be helpful to clinicians, but may also be very attractive to life insurance companies or employers, who may try to assess individual risk levels.
Mr. Chu was lucky to be a good-faith application of technology at the moment when AI was "snapped out" from a random flow of visitors. A new problem arises when a future insurance applicant is marked as a high risk by AI because of a photograph.
In addition, the “visual characteristics of ageing” learned from models may well not be applicable to all colours and facial structures. If the model systematically overestimates the biological age of certain ethnic groups, this error can be repeated and amplified in an extremely hidden manner, even when the miscalculated do not know what happened.
Studies have found that there is a systematic underestimation of the needs of black patients in certain commercial models used to predict the medical needs of patients, since the model uses historical medical expenses as proxy indicators of health status, while black patients have historically received fewer medical resources in the same conditions.
Researchers stressed that FaceAge needed to be tested in a more diverse population before entering clinical practice to identify and eliminate potential errors.
It is much more difficult to deal with these new issues than to train a model.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.